AI agents that do the work — and stop before they do damage.
Software you hand a job to. It works out the steps, does the work, and files a finished result for you to approve. Fifty roles to build from, across finance, marketing, legal, design, content, security and operations.
Builds from $1,900 · managed from $119/mo · you can own every line of it
Where the ceiling sits
Low-risk work runs end to end, unattended — research, drafting, monitoring, reporting, triage, data preparation, analysis. Anything that spends money, sends outbound, publishes, or deploys stops and waits for a person. That is a design decision, not a gap we are working on: an agent that can spend your money while you sleep is a liability, not a feature.
What is an AI agent?
An AI agent is software you hand a job to — it works out the steps, does the work, and hands back a finished result for you to approve.
Agent vs. chatbot vs. automation
| Kind of software | You give it | It decides | You get back |
|---|---|---|---|
| Chatbot | A question, every time | Nothing. It answers what you asked. | A reply in a chat window, which you then have to do something with |
| Automation | A fixed rule you wrote in advance | Nothing. It follows the same path every time. | The same action repeated, until something changes shape and it breaks |
| AI agent | A job, your material, and the limits it must respect | The steps, their order, and when to stop and ask you | Finished work — a drafted reply, a reconciliation, a report — waiting for your approval |
One job, start to finish
A plumbing company takes about forty quote requests a week through its website form. The agent reads each one, pulls out the address and the job type, looks up what comparable jobs were priced at last year, drafts a reply with the number filled in, and has all forty sitting in a review queue before seven in the morning. The owner reads them over coffee, rewrites four, and sends the lot. Nothing left the building without a person looking at it, and nobody spent Tuesday evening writing quotes.
What can an AI agent actually do?
An AI agent can do any digital job with a repeatable shape and a checkable output — reading, drafting, reconciling, researching, monitoring and triaging — using your data, in your business's voice.
Matches last month’s bank statement against your ledger and hands back the lines that don’t tie, with the reason it thinks each one broke.
Drafts this week’s overdue-invoice follow-ups — one per customer, written to get paid without losing the account. You read them and press send.
Reads an incoming agreement and returns a one-page list of the clauses that move liability, IP or indemnity onto you, each with its section number.
Reads six months of reviews and support tickets and returns the handful of complaints that actually repeat, with the quotes behind each one.
Writes search-ad variants for the keywords you are already bidding on, grouped by what the person searching is trying to get done.
Runs the month-end checklist, chases the evidence that is missing, and writes the summary you would otherwise write on a Sunday.
Watches the balance and the burn rate daily and tells you which week you would run short — while there is still time to do something about it.
Turns your inbox, calendar and task list into one morning brief: what is actually due, what is blocked, and drafts for the replies you owe.
Two speeds, split by risk — in your terms:
Reading, drafting, reconciling, researching, monitoring, triaging, preparing data and analysing it all run without you. Sending the email, paying the invoice, publishing the post, pushing the change — those stop and wait. You will see the drafted version and a button. Nothing behind that button happens until you press it.
What is the return on an AI agent?
The return on an AI agent is the loaded cost of the hours it takes off your desk, minus its monthly price and the minutes you spend approving its drafts — for a business handling 1,400 support conversations a month, that works out to about $794 a month, with the setup fee paid back in month 4.
Those four numbers are outputs of the model below, from the inputs named in it — not results we have measured for you. Change the inputs and they change.
The whole calculation, with its inputs named
A 25-person e-commerce brand. 1,400 support conversations a month, 7 minutes each by hand, a customer-service wage of $20.59 an hour.
| Loaded hourly cost | $20.59 × 1.43 | $29.44/hr |
|---|---|---|
| Conversations the agent drafts | 1,400 × 60% | 840 a month |
| Hours reclaimed | 840 × (7 − 2 min) ÷ 60 | 70 hrs |
| Value actually realized | 70 × $29.44 × 0.7 | $1,443/mo |
| Less the plan that covers 840 runs | Team · $649/mo | $794/mo net |
| Setup fee paid back | $2,500 ÷ $794 | Month 4 |
The next plan up (Department, $1,490/mo) covers this volume too, and at these inputs it loses about $47 a month. The bigger plan is the worse plan here. That is the arithmetic we would rather show you now than have you discover in month five.
Where the return actually comes from
Hours reclaimed
The largest and the only one that is simple arithmetic. Minutes per task, minus review minutes, times volume.
Volume absorbed
Work that arrives in spikes stops needing a hiring decision. The agent handles a busy Monday at the same cost as a quiet one.
Errors removed
Mostly the omission kind — the reconciliation line nobody chased, the follow-up nobody sent. The agent does not get bored on line 400.
Coverage outside hours
A quote request that lands at 9pm has a drafted reply waiting at 7am, rather than joining a queue on Monday.
Revenue previously dropped
The hardest to price and the one we leave out of the estimator entirely, because it depends on what you do with the drafts.
What we do not count
Headcount you might not need later, and conversion you might gain. Both are real. Neither is ours to promise, so neither is in the model.
The two assumptions that make it honest
Review time never goes to zero. The agent drafts; you approve. Every number on this page prices 2 minutes of your attention into each task, and the estimator defaults to 3. Any ROI model that assumes nobody checks the work is describing a different product than the one we ship, and its numbers will look better than ours. That is the trade.
A reclaimed hour is not automatically money. We multiply the saving by a realization factor of 0.7 by default. For an owner-operator it is closer to 1.0 — they spend the hour on revenue work instead. For a salaried employee whose headcount is not going to move, it is 0.4 to 0.5, because the hour comes back as slack rather than cash. Set it honestly or the whole model is decoration.
How does Lova build, brand, and deploy an agent?
Lova builds an agent in four steps: scope it against a written test set, build and brand it on your material, deploy it to an account you control, and hand it back reporting into a review queue you approve from.
Scoped against a written test set
Before anything is built we agree, in writing, on the test cases the agent has to pass and the threshold it has to hit. That is how the work gets accepted — and how you check any future change yourself, without taking our word for it.
Built and branded on your material
Your data, your voice, your existing tools. The system prompts, the skill files and the runbook are plain text you can read and edit — not a black box you rent access to.
Deployed to an account you control
Deployed on Cloudflare. On a build, that is your account, your zone, your domain, and your provider keys from day one — we are never in the billing path. On a managed plan we run the copy and you never see a token bill.
Reports back, waits for approval
The agent files its work into a review queue. Low-risk work is already finished. Anything that spends money, sends outbound, publishes or deploys is sitting there waiting for a person — with a two-week parallel-run period in the statement of work, where the agent drafts alongside your team and nothing goes out unreviewed.
The eval set is the part to pay attention to.
Before the build, we both write down the test cases the agent has to pass and the threshold it has to hit — real examples from your business, with the right answer attached. That set is how the work gets accepted, and you keep it. Six months later, when a model changes or someone edits a prompt, you re-run it yourself and see whether anything moved. It is the difference between trusting a vendor and being able to check.
Small tasks, large projects, silos, and coordinated teams
A small recurring task
One agent, one job, on a schedule. Overdue follow-ups every Monday, statements reconciled every month. The Single Task plan exists for exactly this shape of work.
A large project
A leader agent breaks the work down, hands the pieces to specialists, and reassembles the result before you see it. Slower and more expensive per pass — and the only shape that handles work with more than one right answer in it.
Agents working in isolation
Each agent sees only what its own job needs. Nothing is shared between them by default, which is the right default for anything touching customer records or financials.
Agents working as a coordinated team
A leader routes work between specialists and reviews what comes back. It costs more — a task routed through a leader and three specialists is four runs, not one — and it buys you a first-pass review you did not have to do yourself.
Why does a team of narrow agents beat one do-everything agent?
A team of narrow agents beats one do-everything agent because reliability falls as you stack unrelated jobs onto a single agent — the more it is asked to hold at once, the more likely it drifts off the brief or invents something to fill a gap.
You build an agent team the way you build a human one. Somebody runs the project. Somebody runs operations. Somebody coordinates between them. Each person has a job small enough that you can tell, in about ten seconds, whether they did it. That last property is the whole reason it works — for people, and for agents.
A narrow agent has one job, one definition of done, and one test set pointed at it. When the output is wrong you can see which agent produced it and fix that one, and the fix cannot ripple into six unrelated jobs. Stack bookkeeping, ad copy, contract review and customer support onto one agent and none of that is true any more: the tests get vague, the instructions start contradicting each other, and the failure mode stops being "it got the reconciliation wrong" and becomes "it confidently invented something."
So when a vendor offers you one omniscient agent that does everything, they are not offering you more. They are offering you the failure mode, packaged as convenience. We split the work instead, and charge you for the handoffs — a task routed through a leader and three specialists is four runs, not one, and we say so on the invoice.
Who's on an agent team
The catalog splits into 6 leader roles that coordinate and review, and 44 specialists that do one thing each. A leader is where the "one job becomes two to four agents" arithmetic comes from: it does not replace the specialists, it routes between them and reads their work before you do.
The 6 leader roles · typically 2–4 agents each
Everything else in the catalog below is a specialist with one job and one definition of done. You do not need a leader to start — most businesses begin with one specialist and add a leader when there is enough work to route.
Which AI agents can Lova build?
Lova builds from a catalog of 50 agent roles across nine areas of a business — leadership, security, engineering, design, content, finance, legal, marketing and operations — and builds roles that are not on the list when the job calls for it.
How many agents one job needs depends on how granular you want to be. A tight, well-defined task is one agent. A job with judgement in it usually splits into two to four — a leader plus the specialists it routes to. The count on each card below is the typical split, not a minimum you have to buy.
Practically any virtual, digital job can be built and branded to your business, whether or not it appears here. If the shape of the work is different enough from these roles, that is a custom build — same method, scoped to your job.
These are the role definitions from the registry we run in-house, printed as written. A few of them name the specific tools we happen to use; yours get configured against your stack, not ours.
Leadership & coordination
5 roles
Chief of Staff
2–4 agentsCoordinates narrow specialist work, preserves direct synthesis, and final-reviews one integrated leadership-council plan for human build approval.
Coordination
2–4 agentsTurns authorized agenda, inbox, calendar, and task context into a read-only daily coordination brief and editable drafts.
Creativity
2–4 agentsCreates constraint-aware, editable business and creative artifacts from supplied source material.
Clarity
2–4 agentsResearches broad questions and analyzes dense documents with evidence, distinctions, and limitations.
Coaching
2–4 agentsRuns private, repeatable role-play and produces evidence-specific feedback plus a practical prep card.
Security & compliance
7 roles
Security / Guardian
2–4 agentsDefensively reviews digital files, websites and apps, infrastructure, integrations, privacy/compliance, and third-party security without exposing credentials or changing external systems.
Tool & API Permission Review
1–2 agentsReviews supplied tool and API RBAC boundaries and draft-only least-privilege mitigations.
Compliance Officer
1–2 agentsFlags supplied Twilio verification and PII logging evidence plus PostHog telemetry-minimization review needs for legal review.
Security Architecture & Threat Modeling
1–2 agentsMaps supplied Lova-stack assets, trust boundaries, attack paths, blast radius, and approval-gated mitigations without probing or changing systems.
Legal Risk Assessment
1–2 agentsProduces a supplied-evidence threat-model register with attack paths, likelihood, impact, mitigations, owners, and approval gates.
Vendor Check
1–2 agentsReviews supplied vendor and dependency provenance, IaC controls, and permissions without vendor contact or system changes.
Security Operations & Incident Readiness
1–2 agentsReviews supplied detection, audit-evidence, escalation, containment, recovery, and communications plans; returns draft-only readiness gaps and verified-fix checks.
Engineering
5 roles
System Architect
1 agentDesigns scalable Cloudflare-ready architectures and modular APIs.
Docs Fetcher
1 agentExtracts implementation-ready patterns from technical documentation and SDKs.
Skill Creator
1 agentConverts recurring workflows into concise Codex/Claude skill packages.
Webapp Testing
1 agentDesigns destructive end-to-end tests for edge cases, races, and state failures.
Session Handoff Notes
1 agentCompresses session handoffs around decisions, blockers, and next steps.
Design
6 roles
Design Lead
1 agentDesigns accessible, branded React/Next interfaces and user flows.
Taster
1 agentAudits UI/UX against accessibility and modern visual standards.
Frontend Designer
1 agentBuilds componentized, branded, cross-stack interfaces.
Motion Artist
1 agentBuilds performant micro-interactions and transitions with minimal layout shift.
Web Artifacts
1 agentProduces fast single-file prototypes for visual validation.
Brand Guidelines
1 agentEnforces Lova design tokens and brand consistency.
Content
6 roles
Copywriter
1 agentWrites behavior-informed conversion copy with measurable events.
Ghostwriter
1 agentCreates platform-specific social content designed for engagement.
Reel Writer
1 agentScripts short-form video using Hook, Retain, Reward with pacing cues.
Hook Generator
1 agentGenerates opening hooks designed to stop scrolling.
Voice Builder
1 agentKeeps communications aligned with a defined brand voice.
YouTube Thumbnail
1 agentReviews thumbnail concepts for small-size readability and visual CTR.
Finance & bookkeeping
10 roles
CFO
1 agentStructures GAAP-aligned financial statements and liability-aware flows.
Cash Watcher
1 agentMonitors liquidity, burn rate, and lean operating overhead.
Journal Keeper
1 agentConverts transaction data into double-entry entries and flags anomalies.
Debt Chaser
1 agentDrafts relationship-preserving overdue follow-up sequences.
Reconciler
1 agentMatches statements and ledgers and reports discrepancies.
Variance Analysis
1 agentExplains budget-versus-actual deltas and corrective actions.
Margin Analyzer
1 agentAnalyzes unit economics and reducible infrastructure or COGS costs.
Audit Support
1 agentOrganizes financial evidence and logic trails for review.
Tax Prep
1 agentClassifies expenses and surfaces applicable tax considerations.
Close Management
1 agentCoordinates month-end reconciliation and executive summaries.
Legal
3 roles
Contract Review
1 agentReviews agreements for liability, IP, and indemnification risks.
NDA Triage
1 agentChecks NDAs for mutuality, duration, and hidden non-competes.
Signature Request
1 agentPrepares signature-ready handoff packages and tracks status.
Marketing
6 roles
Search & AI-Answer Visibility
1 agentImproves semantic structure, schema, and entity coverage for search and answer engines.
Profile Optimizer
1 agentAligns personal and brand profiles for authority and outbound discovery.
CRO
1 agentDesigns tracked experiments and interaction variants to reduce friction.
Ad Creative
1 agentCreates search-ad variants aligned to audience intent and keywords.
Customer Research
1 agentSynthesizes qualitative feedback into grounded pain-point themes.
Lead Magnets
1 agentDesigns useful low-friction assets that capture intent.
Operations
2 roles
Payroll Planner
1 agentPlans compensation and contractor payout calculations.
Run Campaign
1 agentPlans coordinated go-to-market mechanics and tracking checks.
Not on the list?
The catalog is where we start, not where we stop. Tell us the job and we will tell you whether it is a configuration of something above, a new role built from scratch, or a thing an agent should not be doing at all. We say the third one out loud when it applies.
Can AI agents run marketing and advertising?
AI agents can run the production half of marketing — ad variants, landing-page experiments, social scripts, hooks, search and answer-engine structure, lead magnets and customer research — and they draft all of it, but a person still sets the budget and presses publish.
The expensive part of marketing is rarely the idea. It is the twenty variants nobody has time to write, the review backlog nobody has time to read, and the landing page that never got its second version. 12 of the roles in the catalog exist for exactly that gap: they produce volume at a quality you can edit, so the judgement calls — which audience, which offer, how much to spend — stay with the person who owns the outcome.
What we will not tell you is what it does to your conversion rate or your revenue. Those depend on your offer, your market and what you do with the drafts, and any vendor quoting you a number for them is quoting you a number they cannot control. What we will commit to is the output: the variants get written, the research gets synthesised, the structure gets fixed, and it all arrives ready for you to approve.
Writes behavior-informed conversion copy with measurable events.
Creates platform-specific social content designed for engagement.
Scripts short-form video using Hook, Retain, Reward with pacing cues.
Generates opening hooks designed to stop scrolling.
Keeps communications aligned with a defined brand voice.
Reviews thumbnail concepts for small-size readability and visual CTR.
Improves semantic structure, schema, and entity coverage for search and answer engines.
Aligns personal and brand profiles for authority and outbound discovery.
Designs tracked experiments and interaction variants to reduce friction.
Creates search-ad variants aligned to audience intent and keywords.
Synthesizes qualitative feedback into grounded pain-point themes.
Designs useful low-friction assets that capture intent.
Should I own my agents, or have Lova run them?
Own your agents outright if you want the code, prompts, evals, data and API keys in your own name from day one — fixed-price builds from $1,900. Have Lova run them if you would rather never touch the plumbing — managed plans from $119 a month, by application.
Why these numbers look low
A custom agent build commonly runs $25,000–$80,000. Ours start at $3,500 — because we assemble from a 50-role catalog on a runtime we already built. You are paying for configuration, integration and testing, not from-scratch engineering. Multi-agent orchestration is commonly quoted at $80,000–$200,000; a knowledge agent at $10,000–$70,000. Those are third-party estimates of what this kind of work typically costs, not anyone's published rate card.
Own it outright
You buy the agent, you own it, you operate it, and we train your team to run it. No monthly plan is required — we host nothing on this track, and we are not in your billing path. Best for businesses with someone technical enough to hold the keys, or who simply want no dependency.
1 agent from the catalog, your branding, your data.
- →Email, Sheets and Slack only
- →No custom prompt work
- →You own everything on the checklist below
1 agent with custom prompts and skills, wired into a system you already run.
- →1 real integration — CRM, helpdesk or booking
- →Custom prompts and skill files
- →Written eval set, agreed before the build
1 leader agent plus up to 3 specialists, orchestrated.
- →Up to 3 integrations
- →Leader routes and reviews specialist output
- →Handover session, recorded
Up to 2 leaders and 8 specialists, full orchestration, plus the Security role.
- →Up to 6 integrations
- →Security / Guardian role included
- →Cross-role handoffs and escalation paths
Up to 20 agents across departments, including agents built outside the catalog.
- →Integrations quoted
- →Roles designed for your business, not picked from a list
- →Scoped on a call
What "you own it" concretely means:
- Code —
- The Worker source, config and deploy scripts, in your Git repo.
- Prompts —
- Every system prompt, agent definition and skill file, as plain text you can read and edit.
- Evals —
- The acceptance test set built during the project, so you can verify any future change yourself.
- Data —
- Your knowledge base, source documents and run history, exported in open formats.
- The provider account —
- You hold the API keys, on your own accounts, from day one. Lova is never in the billing path. If you fire us tomorrow, nothing turns off.
- The Cloudflare account —
- Your account, your zone, your domain.
- The handover —
- A written runbook and a recorded handover session.
What is not transferred: the 50-role catalog itself. You get a perpetual licence to the agents we deliver — not the right to resell the catalog they were assembled from. Saying that plainly here beats you finding it in clause 14.
Your own monthly bill on this track, per 500 runs
| Your model choice | Inference | Cloudflare | Monthly floor |
|---|---|---|---|
| Cheapest model tier | ~$2 | $5 | $7–$27 |
| Mid tier | ~$20 | $5 | $25–$45 |
| Frontier tier | ~$60 | $5 | $65–$85 |
| Frontier + prompt caching | ~$33 | $5 | $38–$58 |
Your model choice moves this bill by roughly 15×. We set you up on the cheapest model that passes your eval set, and show you the test results that justify it.
Managed
We host, monitor, tune, secure and support. You never touch the plumbing, never hold an API key, and never see a token bill. Best for businesses that want the output and not the operational surface.
By application · limited onboarding
- →1 agent from the catalog
- →100 runs/mo included
- →$0.45/run over that
- →3-day response
- →1 configured agent, 1 integration
- →400 runs/mo included
- →$0.35/run over that
- →2-day response
- →up to 4 agents — 1 leader + 3 specialists
- →1,200 runs/mo included
- →$0.28/run over that
- →1-day response
- →up to 10 agents
- →3,000 runs/mo included
- →$0.22/run over that
- →4-hr response
- →catalog-wide, cross-department
- →Run allowance negotiated
- →Overage negotiated
- →2-hr + named contact response
Included at every managed tier
Hosting, monitoring, backups, security patching, the approval-queue dashboard, a monthly usage-and-outcome report, all provider costs, and a 12-month price hold. The monthly number is the number.
Overage is capped, not silent
You get an alert at 80% and again at 100% of your included runs. At 120% the agent pauses and we call you. It does not quietly keep spending and show up on the invoice.
What counts as one run
One run is one agent invocation producing one completed deliverable — a draft, a reply, a report or an analysis — including its internal tool calls and retries. A task orchestrated by a leader across three specialists is four runs, not one. Vague units are how vendors end up accused of bad faith, so that is the definition, in writing, before you sign anything.
Care plans
Optional, and only on the track where you own the agent. Industry guidance puts ongoing maintenance at 15–30% of the build price a year; these plans are that number, itemised. Maintenance is the product here, not an upsell bolted on at the end.
Watch
Know it is up, and know what it costs.
- ✓Uptime and error monitoring
- ✓Monthly usage-and-cost report, so the provider bill never surprises you
- ✓Model-deprecation alerts
- ✓Email support · 2-day response
Keep Running
The one that survives a model being retired under you.
- ✓Everything in Watch
- ✓Runtime security patches
- ✓One model-deprecation migration per 12 months, handled
- ✓Priority bug fixes · 1-day response
- ✓10% off add-ons
Tune
For agents whose output quality is worth watching monthly.
- ✓Everything in Keep Running
- ✓1.5 hrs/mo tuning and eval work, rolling over to 3
- ✓Monthly output-quality review against your eval set
- ✓Quarterly strategy call
- ✓20% off add-ons · 4-hr response
Build + 12 months of care · 1 month free
The costs that usually show up on an invoice
Published here instead, because finding them later is what turns a good project into a dispute.
- Integration
- A published rate card, fixed and quoted before you sign: simple $600 · standard $1,500 · complex $3,000. Each build tier includes a stated number; the next one is a new fixed quote.
- Data preparation
- A separate, quoted readiness step before the build. If the data is not ready we quote the cleanup or decline the project. It never gets absorbed into a fixed build price and discovered as a delay.
- The provider bill
- Modelled at signing across three model choices, with the eval results that justify the cheapest one that passes. On a managed plan it is included and you never see it.
- Ongoing maintenance
- Industry guidance is 15–30% of the build price per year. On a $3,500 build that is $525–$1,050 a year — the care plans below run $588–$2,268. That is the same range, said out loud.
- Human supervision
- Budget 2–5 minutes per draft. It is in the estimator on this page. Most vendors leave it out, because their model assumes you never check the work.
- Model deprecation
- Keep Running and every managed tier include one migration per 12 months. Beyond that it is quoted.
- Compliance adders
- Healthcare +$4,500 · finance +$3,500 · legal +$1,500.
- Training and change management
- Included in every build, with a two-week parallel-run period written into the statement of work.
Discounts that carry over: SMB and veteran 10% · annual prepay 15% · nonprofit 20% or 15% · referral $250 each way · a second project within six months, $1,000 off. Stacking is capped at 20% in total. No hourly rates, ever — out-of-scope work becomes a new fixed quote.
What we will actually put in writing
- →Acceptance against a written eval set — we agree the test cases and the pass threshold before the build. "It passes the test we both signed" is objective and you can re-run it. An accuracy percentage on open-ended output is not, so we do not quote one.
- →Track A: 30-day acceptance. If it fails the agreed eval set after two remediation passes, we refund the build less a 25% discovery fee.
- →Track B: cancel any time in the first 60 days and keep your exports. The setup fee is non-refundable.
- →Runtime uptime 99.5%/month on managed plans. The remedy is a service credit, capped at that month’s fee.
- →Data portability — full export in open formats within 10 business days, on any plan, at any time, at no charge.
- →No silent model substitution. Changing the model behind your agent requires your prior written approval of the named alternate.
- →A 12-month price hold on every managed tier.
- →A delivery date, with a stated credit if we are the reason it slips.
What is missing from that list is deliberate. No accuracy percentage on open-ended output, because it is unfalsifiable. No promise about your revenue, conversion, close rate, headcount or hours saved, because those depend on decisions on your side. Nothing implying an agent acts unattended on money, outbound, publishing or deploys, because it does not. Compliance is a property of a deployment, not a product, so we will say which named controls we build to and nothing broader. See how we handle data.
Switching later
The two tracks are a ladder, not a trap. You can move in either direction, and you can leave without moving at all.
Own it, now run it for me
The migration fee is one month of the target tier, flat. No setup fee. You keep everything on the ownership checklist; we operate a copy.
Renting, now I want to own
The buyout is the build tier price, minus 50% of what you paid us in subscription over the last 12 months, capped at 50% of the build price. Setup fees already paid credit in full.
Exit without buying anything
Always available, always free. On any plan, at any time, we export your prompts, eval sets, knowledge base and run history in open formats within 10 business days at no charge.
What would an agent save my business?
What an agent saves your business is the loaded cost of the hours it drafts away, minus the plan fee and the minutes you spend approving — put your own numbers in and this estimator returns the cheapest managed plan that actually covers your volume, or tells you that none of them earn their fee.
At 840 runs a month, the cheapest plan that covers you is Team at $649/mo. After that fee you keep about $794 a month. The $2,500 setup fee pays for itself in month 4.
Department ($1,490/mo) covers this volume too, and nets −$47 a month. The bigger plan is the worse plan here, so we do not recommend it.
- Loaded hourly cost
- $29.44/hr
- Tasks the agent drafts
- 840 of them
- Hours reclaimed
- 70 hrs/mo
- Value realized
- $1,443/mo
- Recommended plan
- Team — $649/mo
- Net after the fee
- +$794/mo
- Cost per task, agent vs you
- $1.75 vs $3.44
- Setup paid back
- Month 4
This is a model, not a forecast, and it runs against managed-plan pricing where all provider costs are included. It prices your review time in, and it multiplies the saving by a realization factor rather than assuming every reclaimed hour turns into cash. If the answer comes back negative, that is the answer — we would rather lose the sale than sell you a plan you will cancel in month five.
Frequently asked
What happens when an AI agent gets something wrong? +
You catch it in the review queue, because the agent hands you a draft rather than a finished action. That is the whole reason the approval step exists: every system built on a language model gets things wrong sometimes, and any vendor telling you theirs does not is either not measuring or not telling you. What we can commit to is the eval set — an agreed list of test cases with a pass threshold, run before delivery and re-runnable by you after any change. When something slips past it, the fix is a new test case plus a prompt or skill change, and on a care plan that is included.
What does an AI agent cost to run every month, on top of the build? +
On a build you own, the floor is roughly $7 to $85 a month at 500 runs, depending entirely on which model you choose — about a 15× spread — plus $5 of Cloudflare. We set you up on the cheapest model that passes your eval set and show you the test results that justify it. On a managed plan there is no separate bill at all: hosting, monitoring and all provider costs are inside the monthly price, so a $119 plan costs $119.
What happens to my agents if I stop paying Lova? +
On a build you own, nothing turns off. The code, prompts, evals and data are already in your accounts, the API keys are already in your name, and we were never in the billing path — cancelling a care plan stops the monitoring and the patches, not the agent. On a managed plan the hosted copy stops, and we export your prompts, eval sets, knowledge base and run history in open formats within 10 business days at no charge. That export is free on every plan, at any time, whether or not you are leaving.
Do I need to be technical to use an AI agent? +
No. What you need is to be able to tell good work from bad work in your own business, which you already can. Day to day you open a review queue, read what the agent drafted, edit what needs editing, and approve. The technical surface — prompts, deployment, keys, model choice — is either handled on a managed plan or handed to you with a written runbook and a recorded walkthrough on a build. Deciding you would rather not run it yourself is a normal outcome, and the reason both tracks exist.
How is this different from just using ChatGPT? +
A chat window answers you and then forgets. An agent is set up once with a job, your data, and access to the systems the job touches, then runs on its own schedule and files its output where you work. Practically, the difference is that nobody has to remember to open it, paste the context in, and check the format — the reconciliation runs on the first of the month whether or not anyone thought about it. A chat subscription is cheaper. It is also a tool your team has to drive; this is work that arrives finished.
Can an AI agent send emails or spend money on its own? +
No — and not because we have not got around to it. Low-risk work runs end to end unattended: research, drafting, monitoring, reporting, triage, data preparation and analysis. Anything that spends money, sends outbound, publishes, or deploys stops and waits for a person to approve it. An agent that can move money or email your customer list while you sleep is a liability sitting on your balance sheet, not a feature, and we do not ship one.
What if the agent does not actually work? +
On a build, you have 30 days from delivery. If the agent fails the eval set we both signed and two remediation passes do not fix it, we refund the build less a 25% discovery fee. On a managed plan you can cancel inside the first 60 days and keep every export; the setup fee is not refunded. What we will not offer is money back if you do not see a return, because that would mean underwriting decisions on your side of the wall — whether the reclaimed hours get used for anything, whether the drafts get reviewed. The estimator on this page is deliberately conservative for the same reason.
How long before an agent is actually running? +
One week for an off-the-shelf catalog agent, two weeks for one built around your prompts and a real integration, four weeks for a leader plus three specialists, and six to eight for a full department. A managed Single Task plan is faster still, because there is nothing to build. Add the two-week parallel-run period after delivery, where the agent drafts alongside whoever does the job today and nothing goes out unreviewed, before you count on it.
Built right. Kept running.
Tell us the job you would hand to an agent and we will tell you whether it is a $1,900 catalog agent, a $3,500 build, or something we would talk you out of.