AI for sales teams without made-up numbers

AI makes up numbers when it has no source to read. The fix is to give it sources and rules: standing instructions, a memory you can inspect, a verification standard, and a person who approves anything that leaves the building. I run my sales operation this way, and this page lays out the method.

Why does AI make up numbers?

An AI assistant writes the text that is most likely to fit, and it does not look up facts unless you hand it something to look at. Ask for a close rate it has never seen and it can still produce one that sounds right. A gap in the input becomes a plausible figure in the output.

This is how these tools write, so it shows up in every assistant. Switching assistants does not change this. Removing the gaps and checking the output does.

Four controls do that work:

1. Standing instructions
2. Memory the assistant can read
3. A verification standard
4. Human approval before anything ships

The sections below cover each one.

What are standing instructions, and what belongs in them?

Standing instructions are a written manual the assistant loads at the start of every session. Mine is a versioned file called CLAUDE.md. Every session loads it, and every deliverable follows it.

Put four things in it:

• Voice and brand rules. How you write, which words you never use, and how your documents look.
• Verification standard. How figures get sourced and checked. More on this below.
• Tool routing. Which tool handles which job, so the assistant does not improvise.
• An approval list. What never ships without a person's sign-off.

Rules work best when something enforces them, such as a template or a check that runs when a document is built. A slip caught there never reaches a reader.

Keep the manual under version control so you can see what changed and when. When the assistant gets something wrong, add a line to the manual. The next session starts with the fix.

How do you give an assistant memory that stays accurate?

Keep memory in a structured vault of plain notes that a person can open and correct. Mine is an Obsidian vault. It indexes brand standards, playbook positioning, people and how they decide, recurring plays, and a running log of every deliverable produced.

A few habits keep it accurate:

• Use readable files. If you cannot read what the assistant remembers, you cannot correct it.
• Give each subject one note. A fact that lives in one place gets fixed in one place.
• Log every deliverable. Record what was made and which sources it used. Next month, you can check an old number in minutes.
• Separate facts from guesses. Mark each line as confirmed or unconfirmed. An assistant that reads a guess as a fact will repeat it with confidence.

Review the vault now and then. Memory can hold a wrong note as easily as a right one.

How do you make the assistant improve over time?

Schedule a review at the end of the day and a load step at the start of the next one. My system runs both. A scheduled evening reflection reviews the day's sessions, extracts lessons that apply beyond one task, and writes them into memory. A wake routine loads those lessons back before the next session.

The word that matters is "generalizable." A lesson about one prospect's name helps once. A lesson about how to check a source helps every day.

Read what the reflection writes, at least at first. A bad lesson saved to memory will shape every session after it, so treat the nightly write like any other output that needs a check.

What is a verification standard, concretely?

A verification standard is a short list of rules that decide whether a number is allowed to ship. Mine has four:

1. Every figure has a source. Name the CRM report, billing export, rate sheet, or document it came from. If you cannot name one, the number does not ship.
2. Unconfirmed items get flagged. The assistant writes "unconfirmed" and says what is missing. It does not fill the gap with a guess.
3. Money gets recomputed independently. A second calculation, run a different way, checks the first. Financial models get this before they go out.
4. A person approves anything external. Client-facing material waits for a human decision.

Here is rule three at work. My commission engine includes an independent Python recompute that checks every figure before a statement goes out. The audit found and fixed real silent errors, including a formula desync that had been dropping dollars.

Rule two shows up in my live sales dashboard. When data fails to load, it shows an error banner. A zero that looks plausible would pass without anyone asking questions.

Which sales work is a good first project?

Pick work where the inputs are documented and the answer can be checked. These four fit:

• Recurring reports. The same report on a schedule, with sources named. Mine is a weekly pipeline report that arrives every Friday morning.
• Data cleanup with an audit trail. Every change is logged, so you can undo or explain it.
• Calculators built on documented rates. The inputs are written down, so the output can be traced. My funnel calculator uses documented stage-to-stage conversion rates.
• Compensation checks. Money work, with an independent recompute.

The CRM clean-up is a good template. A full read of 25,150 company records in our CRM found 454 duplicate pairs and the pairs were tiered by confidence. A script merged the 183 high-confidence pairs. Each merge carried a preservation note for any data the merge would have lost, plus a full audit trail. The rest went to a reviewed queue, where a person decides.

That job worked for four reasons. The data was in the CRM, so nothing had to be guessed. A script did the repetitive part. Every change left a record. Uncertain cases went to a person.

Avoid starting with open-ended work like "forecast next year" or "tell me what the market will do." Those questions have no source to check.

What should stay human?

Approvals, client communication, pricing judgment, and the final read of any number that goes outside stay with a person. The assistant drafts, and you decide.

• Approvals. Your approval list names what cannot ship on the assistant's say-so. A person signs off every time.
• Client communication. A client email carries your name and your relationship. Read it and send it yourself.
• Pricing judgment. The assistant can recompute a model and show the arithmetic. Whether a price fits a client and a deal is your call.

The assistant handles volume: reading records and running the math twice. You handle the decisions that carry risk.

Where should you start?

Pick one recurring report. Write down where every number in it comes from, and mark any number you cannot trace. That list is the first draft of your verification standard. Put it in a standing instructions file, and have a person approve the next report before it goes out.

If you want to talk through your setup, use the Connect link at the top of this page.

About the author. Steven Bartz is Director of Sales at Five Nines Technology Group in Lincoln, Nebraska. He runs his sales operation on an AI system built around one rule: every figure traces to a source.