An API changes. A model version gets retired by the provider. Accuracy drifts as your inputs shift. None of it announces itself, the system keeps running and quietly starts being wrong. We monitor, tune, update, and fix the systems we build, and the ones we didn't, for a fixed monthly fee with a named engineer attached.
From $249/month · Month-to-month, no lock-in · Documentation handed over whether you stay or go
System status
Support agent — chatbot
Checked 4 minutes ago
Invoice pipeline
Checked 12 minutes ago
CRM integration
Schema change handled — no action needed
Customer portal
Checked 2 minutes ago
Lead-scoring model
Checked 1 hour ago
Illustrative example, not a live client dashboard.
Trusted by teams in the US, UK, and India
Siddhraj
Unoloft
Kofeko
3nStar
Veda
Cerata
Shubham
Consultup IndiaSiddhrajUnoloftKofeko3nStarVedaCerataShubhamConsultup IndiaIf you leave a well-built web app alone, it will probably still work next year. That intuition is reasonable, it's why maintenance plans feel like an upsell, and it does not transfer to AI systems.
Every major provider retires model versions on a published schedule. When yours is retired, your system stops working on a date somebody else chose. Moving to the replacement isn't a switch, outputs shift, prompts behave differently, and the whole thing needs re-testing against your actual cases. This is a calendar event, not a risk, and it's exactly the kind of thing a maintenance plan exists to track.
The model was tuned on the documents, tickets, and products you had at launch. You've since added a product line, changed your pricing page, and started serving a different customer. Nothing broke. The answers just got worse, gradually, in a way nobody notices until a customer points it out. This is what a monitoring plan is built to catch, regular review against real outcomes, not a one-time launch check.
Your CRM ships an update. A vendor changes an API response. A source adds a required field. Integrations are the most common failure point in any AI system and they fail on somebody else's release schedule, not yours.
A crashed server pages someone. A drifted model returns a confident, plausible, wrong answer and logs a success. Without monitoring built specifically for it, the failure mode of an AI system is looking fine.
This is roughly what a good month looks like on a plan. None of it would have reached you.
A source system added a required field. The connector was updated before the next scheduled run.
An API key was approaching its rotation deadline. Rotated ahead of time, nothing interrupted.
The provider announced a deprecation date for a model version you're running. Diarised, migration scheduled with room to spare.
Three dependency updates applied, one of them a security patch.
A load failed overnight on a rate limit. Retried automatically, then the retry window was widened so it wouldn't recur.
Someone on your team asked for an extra field on a report. Done inside the plan.
Monthly note sent: what ran, what was fixed, what's coming, and what we'd suggest next.
Nothing on that list would have been an emergency this month. Two or three of them would have become one within a quarter, usually noticed by a customer, or by a number that was wrong for longer than anyone realised.
You're not buying incident response. You're buying the month where the incident didn't happen.
Monitored: Response latency, escalation and fallback rates, cost per conversation, retrieval quality where a knowledge base is involved, and periodic review of answers against real conversations.
Typically needs attention: Prompt tuning as your content changes, re-indexing when documents are added or revised, help migrating to a replacement when a model version is retired, and tightening a fallback that fires too often or not often enough.
If we haven't built it yet, that's AI agent development or RAG chatbot development, this covers what happens after launch.
Monitored: Run success rates, queue depth, error patterns, third-party API health and deprecation notices.
Typically needs attention: Connector updates when a vendor changes an endpoint, retry and rate-limit tuning, error-queue review so failed items get resolved instead of accumulating, credential and token rotation.
If we haven't built it yet, that's workflow automation, this covers what happens after launch.
Monitored: Freshness against expected arrival windows, row-count anomalies, schema drift, warehouse and connector spend.
Typically needs attention: Schema changes at source, backfills after an outage or a correction, query and cost optimisation, adding sources as they appear.
If we haven't built it yet, that's data warehousing and ETL, this covers what happens after launch.
Monitored: Uptime, error rates, performance, dependency and security advisories, backup integrity.
Typically needs attention: Dependency and framework updates, security patches, small feature requests, and the ordinary accumulation of “can it also do this.”
A vendor changes a response field. Your integration keeps running and starts writing blanks.
Reports look slightly off. Someone assumes it's a reporting bug.
A client asks why a number is wrong. Nobody can say how long it's been wrong.
Your original developer is on another project, or unreachable, or gone. Finding someone who can read the code takes days.
Fixed. Then the real work starts: identifying every record affected and correcting a month of downstream data.
Cost: the fix, the backfill, the client conversation, and a quiet loss of confidence in every other number the system produces.
The field-level check fails on the first run after the change. Alert reaches us, not you.
Connector updated. Affected rows re-processed from the last good run.
Logged.
It appears as one line in your monthly note.
Cost: included.
Your developer went quiet. The agency that built it doesn't do maintenance. The person who owned it left and took the context with them. The automation has been broken for two weeks and nobody can find where it runs. That's a normal Monday and it's a normal engagement. Here's how it goes.
A two-week audit, first, $349. We map what exists, what it depends on, where it runs, what's undocumented, and what's actually broken versus merely alarming. You get a written assessment and a recommendation. It's a fixed price and it's yours to keep whether or not you continue with us.
Then an honest verdict. Sometimes the answer is a rebuild. More often it isn't. The engineering is usually sound and what's missing is everything around it, no monitoring, no documentation, no version control, no owner. Wrapping a working system in those four things is a fraction of the cost of replacing it, and we'll tell you when that's the case even though the rebuild is the bigger invoice.
Then we take the pager. Once the audit is done and gaps are closed, it moves onto a standard plan and the alerts route to us.
What we'll need: repository access or the code itself, credentials or a path to reissuing them, and hosting access. If nobody has any of those, say so on the call, it's recoverable more often than you'd expect, but it changes the shape of the first two weeks and we'd rather plan for it than discover it.
Book an audit call$249/month
Monitoring and alerting, uptime and error tracking, dependency and security updates, failure response, and a monthly written note.
Best for: one system, stable, low change rate.
$599/month
Everything in Essentials, plus periodic accuracy review with tuning where needed, help migrating when a model version is retired, connector updates when vendors change APIs, backfills and error-queue review, and a day or two of small feature work each month. A named engineer.
Best for: most clients, most systems.
$1,499/month
Everything in Standard, plus a larger monthly allowance, priority response, quarterly review of the whole stack including cost, and a roadmap conversation rather than a ticket queue.
Best for: several systems, or one that's central to how the business runs.
Partner's "priority response" means priority queuing within these same commitments, not a different published number, we'd rather commit to a number we can honour on every tier.
| Essentials | Standard | Partner | |
|---|---|---|---|
| System down | 4 hours | 4 hours | 4 hours |
| Degraded or producing wrong output | Next business day | Next business day | Next business day |
| Everything else | 2 business days | 2 business days | 2 business days |
| Overlap hours with US Eastern and UK | 2–3 hours daily | 2–3 hours daily | 2–3 hours daily |
Month to month. Thirty days' notice, no minimum term, no exit fee. Documentation, code, and credentials are yours throughout and handed over cleanly whether you're leaving us or just taking it in-house. A plan you can cancel is the only kind worth selling, if we're not earning it, we'd rather you stopped paying than stayed locked in.
Being specific about this up front is how the plan survives contact with year two.
Included
Monitoring and alerting · failure diagnosis and fixes · periodic accuracy review and tuning where needed · help migrating when a model version is retired · connector and integration updates when third parties change · dependency and security updates · backfills and reprocessing · small feature work up to the monthly allowance · a named engineer and a monthly written note.
Not included
New systems or new integrations · significant new features beyond the allowance · redesigns · data-entry or manual operational work · your third-party licences, API usage, or hosting costs, which stay in your name and are billed to you directly · support for systems we haven't audited.
On the last two, plainly: we don't resell infrastructure or model usage, you pay providers directly at their prices, so there's no margin for us in your usage going up. And we won't take responsibility for a system we haven't looked at, the audit above exists precisely so we're never guessing about what we've agreed to keep alive.
A three-step Zapier automation between two mainstream tools. If it breaks, it breaks loudly and someone can fix it in an afternoon. Pay for that afternoon when it happens.
A static site or a system with no AI, no integrations, and no data flowing through it. Very little decays. Check the dependencies twice a year.
A system your own team already owns properly. With monitoring, documentation, and someone whose job it is. That's the right end state. If you're there, we're overhead.
The plans exist for systems with moving parts. If yours doesn't have those, we'll say so on the call.
No. You own the code, models, and documentation outright from final payment. The plan buys monitoring and maintenance, not access. Cancel any time with thirty days' notice and everything stays yours, documented.
Yes, it's roughly a third of our maintenance work. It starts with a fixed-price two-week audit so we both know what we're taking on. The written assessment is yours regardless of what you decide next.
We track announced retirement dates for every model in your systems. Before the deadline we migrate to the replacement, re-test against your real cases, adjust prompts where behaviour has shifted, and check outputs against real cases before calling it done. It's included on Standard and Partner plans.
Traditional maintenance is mostly about keeping something running. AI maintenance adds keeping it correct, accuracy review, tuning, and model migrations. A system can be running perfectly and producing worse answers than it did in March, and only the second kind of attention catches that.
A named engineer who knows your setup, reachable on email and a shared Slack or WhatsApp channel. Response commitments are in the table above and they're in the contract, not just on this page.
We commit to 2–3 hours of daily overlap with US Eastern and UK working hours, with a same-business-day response on anything urgent.
Then you've had the outcome you're paying for, and the monthly note will be short. If a system genuinely proves stable enough not to need the plan, we'll suggest moving you down a tier. It's happened.
No. Provider and hosting accounts stay in your name and you're billed by them directly at their prices. We don't resell usage and don't take a margin on it, which is why we can tell you when a cheaper model or a smaller instance would do the same job.
Book a 30-minute call. Whether we built it or someone else did, we'll tell you what's likely to break first, what it would take to keep it healthy, and what that costs a month.
