AI for MSP QBRs: what to automate and what to keep human

AI for MSP QBRs works best when it does the assembly and leaves the judgment to you. Pulling ticket trends, reading backup results, drafting the narrative and checking it for contradictions are jobs a model does well and fast. Deciding what a number is, what you recommend, what it costs and what the customer agreed to are jobs that stay with a person. This guide sorts the work of a business review into those two piles and lists the guardrails that keep an AI-written review safe to put in front of a customer.

Why the line matters

A review carries your name. A model can read a trend the wrong way round, draw the wrong conclusion from a correct number, or phrase something in a way you would never say to that customer. One wrong figure on a scorecard costs more trust than the hours the AI saved. So the question is not whether to use AI for business reviews; it is which parts of the review a model is allowed to produce on its own, and which parts it may only draft for you to approve.

Outside the MSP market the same rule shows up in general QBR tooling. A 2026 roundup from Cotera puts it as "the agent prepares; the CSM decides", and argues that the software should keep evidence separate from interpretation, because "weekly active users fell 18%" is a fact and the reason behind it is a claim that needs support.

What to automate

1. Gathering the evidence

The slowest part of QBR prep is exporting from five tools: the PSA for tickets and CSAT, the RMM for patching and device age, the backup console, the security stack, the lifecycle tool for warranty. This is pure retrieval and a machine should do all of it, for the right customer and the right period, every time.

2. Turning raw records into fact sheets

A model is good at reading 300 tickets and summarizing what people asked for, which issues your monitoring caught on its own, and which themes repeated. It is also good at stripping things that should never reach a page: anything that looks like a password, a key or a personal phone number.

3. The first draft of the words

Narrative is where AI saves the most time. Writing chapter by chapter, with a storyline planned first, produces a review that reads as one piece: each chapter knows what the earlier ones said, and a fact is emphasized once rather than repeated in every section.

4. Consistency checks

After the draft, a read-back pass can catch a number that came out differently in two places, a claim that contradicts another, or a phrase used three times. This is tedious for a person and reliable for a model, provided it returns targeted corrections instead of rewriting the pack.

5. Finding benchmarks, with sources

"Is that normal?" is a fair customer question. AI can look up a published figure for a firm of the same industry and size, as long as the figure prints with its source and year, and a figure with no source never prints.

6. Roadmap suggestions and meeting prep

Spotting hardware past warranty, a backup that has failed for three days, an MFA gap or a ticket theme, and proposing it as a roadmap item, is pattern matching across records. So is a prep sheet listing what changed since the last review, what to ask, and which facts have gone stale. See how to build an IT roadmap for MSP clients for how those suggestions fit the plan.

What to keep human

The numbers themselves

A figure on a review should be computed from the record, never written by a model. If a model is allowed to type numbers, sooner or later it types one the evidence does not contain. The safe pattern is the reverse: compute every figure first, hand the model only those figures, and drop any number in its output that the evidence does not already print.

Money

Your MRR, contract amounts, costs and margin are not inputs a model needs to write about a quarter, and sending them out is a risk you do not have to take. Keep the money pages deterministic and priced from your own records, and keep prices out of the prompts.

Recommendations and priorities

AI can propose that a customer refresh 14 laptops in Q2. Whether that goes ahead of the network project, given what you know about the owner's plans for next year, is an account manager's call. Treat every AI proposal as a draft line you accept, edit or drop.

Tone and what to leave out

You know that this customer's office manager took last quarter's outage personally, and that the CFO reads only the first page. Read every chapter before the meeting, rewrite the ones that land wrong, and hide pages that do not belong in this room.

Decisions

Approve, defer and decline belong to the customer, recorded by name and date. An AI summary of "what was agreed" is not a record of what was agreed. Capture decisions directly, in the meeting, against the item they apply to.

Guardrails to look for in any AI QBR tool

• Computed numbers. The tool should compute figures from your data and stop the model inventing new ones.
• Missing data drops out. A section with no data should disappear, not print a zero or an empty chart.
• Evidence on every claim. Each risk or recommendation should link to the device, ticket or contract behind it.
• No money in prompts. Ask the vendor what leaves your account when the AI runs.
• Read as you. An AI answer about a customer should only draw on records your own permissions can open.
• A review step. Chapter-level rewrite, line edits and the ability to hide a page before anything is sent.
• Diagrams built from evidence. A chart the model designs should only use values already in the record.

How Current splits the work

Current's Strategic Business Review follows this split. A run reads every record your connected services hold for the customer and the period, builds a fact sheet per service (redacting anything that looks like a password, key or personal phone number), plans the storyline, writes one chapter per step, sources benchmarks, and finishes with a read-back pass that returns targeted corrections. It takes about six minutes and keeps going if you close the tab.

• Numbers are computed and words are written: the writer may only repeat a figure the evidence already prints.
• No money reaches a prompt: MRR, contract amounts, cost and margin stay on deterministic pages, and open opportunities reach the writer by name and stage, without amounts.
• An explainer diagram whose values are not already in the evidence is dropped rather than drawn.
• A chapter with nothing to report drops, and the agenda renumbers.
• Benchmarks print with their source and year, and a figure published for a different industry is never used.
• Check and adjust is its own stage: rewrite one chapter in seconds, edit any sentence on the page, move the bold, or hide a page.

On the roadmap, Draft a plan proposes goals, initiatives and sequencing for one company, and nothing is written until you accept each line. Ask the plan answers questions from the company's records, read with your permissions, and when nothing answers the question it says what it checked. In both, prices, contract amounts, quote totals and your costs are removed before anything is sent, and values are attached in your browser afterwards. Decisions are stamped by a person, in the meeting, with who decided.

AI can make mistakes, and Section 10 of the Terms sets out that AI output is a draft for your review. AI is included in Current's per-user price (see pricing).

Example: one backup chapter, two piles

Take the backup chapter for a fictional 40-person engineering firm, Ridgeway Design. The machine side of the work: read every protected system's last result for the quarter, count the verified restores, pick the dated boot-verification screenshots, and note that two systems went three days without a completed backup in August. The writing side: a draft paragraph saying that both gaps were caught by monitoring and closed the same week, and that a recovery drill is due next quarter.

The human side: confirm that "closed the same week" is what happened, decide whether to mention that one of the two systems is the managing partner's laptop, and decide whether the drill is a recommendation or already approved. Five minutes of reading, and the chapter is safe to present.

A simple rollout plan

• Pick three customers with clean data in your PSA and RMM and run AI-drafted reviews for them first.
• Have the account manager read every chapter and note each edit they made and why.
• If the same correction keeps appearing, fix the data at the source (a mis-mapped company, a missing feed) rather than editing around it each quarter.
• Expand to the rest of your top-tier customers once a draft needs only light edits.
• Use the saved prep hours on the part AI cannot do: the conversation about what the customer should do next.

For the meeting structure itself, see how to run an MSP QBR. If you are comparing tools, see QBR software for MSPs.