Most MSPs underprice AI retainers because they price the software and forget the humans who mop up after it. The bot looks cheap on a vendor quote. The after-hours Slack thread where your senior tech explains why the bot reset the wrong shared mailbox is not on that quote. If you sell a flat AI add-on the way you sell antivirus, you will fund the client's experiment with your margin.
A retainer can work. It just has to look like managed service math, not SaaS markup cosplay. You are selling response quality, scoped change, and a clear line between included work and billable project work. Get those three wrong and the retainer becomes a support sink with a fancy name.
Price the failure modes, not the model
Start with a month of ticket samples from one vertical you already serve well. Pull fifty tickets from a clinic, a machine shop, or a property manager. Mark which ones an AI tier-1 agent could close without a human, which ones need a human in the loop, and which ones should never touch a model because the blast radius is ugly. Password resets with clean MFA flows might be closable. Permission changes on a shared drive that holds payroll files are not. Invoice-matching helpers for an accounts payable clerk sit somewhere in the middle and usually need review.
That tagging exercise gives you a close-rate guess. Be conservative. If you think the bot can close thirty percent of tier-1 volume, price as if it closes twenty. The gap is where your techs live. Every false close creates a second ticket with an angry user and a longer handle time. I have watched a white-label email triage bot "resolve" vendor invoices by filing them under the wrong entity. The AP manager only noticed at month close. Your retainer ate eight hours of cleanup that nobody budgeted.
Build the monthly number from three buckets. Bucket one is platform cost: seats, tokens, hosting, and any white-label fee you pay upstream. Bucket two is included human time: a fixed number of hours for prompt tweaks, knowledge base updates, and review of escalations. Bucket three is contingency for the ugly week when the client dumps three years of PDFs into the knowledge base and retrieval quality tanks. If bucket two is zero because you told yourself the AI is self-serve, you are lying to your future self.
A practical pattern for a fifty-user professional services firm: charge a base AI retainer of $800 to $1,500 a month on top of the existing managed stack, include two hours of knowledge maintenance, and set a hard cap on automated actions that touch identity or finance systems. Anything above the two hours bills at your project rate. Token overages bill as a pass-through with a small admin fee, itemized so the client sees usage. When they ask why the bill moved, you show the report instead of arguing vibes.
Do not bury unlimited prompt engineering in the retainer. Clients will ask you to "just make it sound more like our brand" every week. That is creative services. Put a change window in the agreement: two scheduled tune-ups a month, requests logged in the PSA, and a freeze during their busy season if you support accountants or nonprofits around filing deadlines. Unlimited polish is how a $1,200 retainer becomes a part-time job for your best engineer.
Write the contract like you expect an incident
Spell out what the AI is allowed to do without approval. Read-only answers from a vetted knowledge base are one thing. Creating tickets is usually fine. Resetting passwords, modifying groups, sending mail as a user, or changing firewall rules should require a human confirm step or stay off the table. Put that list in the SOW in plain language. When a vendor ships a new "agentic" feature mid-contract, you need a clause that new action types stay disabled until you and the client sign off. Otherwise the product updates itself into a liability and your retainer silently expands.
Define success metrics that match service desk reality. Average handle time and first-contact resolution matter more than "engagement" or daily active users. A bot that chats a lot and closes little is a toy. Agree on a quarterly review where you look at deflection that stuck, escalations with bad context, and tickets the bot should never have touched. If deflection is weak after ninety days, you want a contractual path to shrink scope or cancel without a fight. Clients respect an MSP that exits a bad pilot early more than one that keeps billing while everyone pretends the dashboard looks fine.
Staff the retainer on purpose. Name a primary owner on your side, even if it is ten hours a month. Rotating "whoever is free" ownership is how knowledge base drift happens. The owner updates sources when the client changes VPN vendors, retires an old line-of-business app, or hires a seasonal crew that needs a different onboarding script. Without that owner, the bot answers yesterday's environment with today's confidence.
Watch the support cost weekly for the first two months. Track tech minutes spent on AI-related tickets separately in your PSA. If those minutes climb while the retainer stays flat, raise the price at renewal or cut features. I have seen shops freeze the AI add-on price for a year out of fear of looking expensive next to a competitor's $49 seat pitch. That competitor is often selling a thin wrapper with no PSA integration and no after-hours coverage. You are selling a service desk with a model attached. Price the desk.
Package the retainer so sales can explain it in one breath. Example: "AI helpdesk layer for your existing users, includes knowledge updates twice a month, human review on identity changes, monthly usage report, and a ninety-day exit if it does not reduce tier-1 load." That sentence is dull on purpose. Dull scopes survive QBR season. Flashy scopes generate surprise invoices and awkward calls with the person who signed the order form.
If a prospect wants "unlimited AI" for a number that barely covers the vendor bill, decline. Point them to a fixed pilot with a start date, end date, and a written success bar. Pilots protect both sides. Retainers should come after you have seen real ticket mix, not after a lunch-and-learn where the bot recited the employee handbook. Your margin is not a demo environment. Treat it that way and the AI line of business stops being the department that funds everyone else's innovation theater.