Where should a small business start?

Start with one job that repeats every week and where experience changes the result, such as quoting a repair, triaging appointment requests or approving an exception to a policy. Walk through a recent real case with the person who does it best and write down what they check and why.

Does a Map of Work stop AI vendors from seeing company data?

No. Anything sent to an outside AI service has been disclosed to that provider. A Map of Work reduces dependence on one vendor and lets each tool receive only the context a task needs, but over many requests a provider can still learn a lot about how you work.

Do you need special software to build a Map of Work?

No. A first map fits in a shared document or spreadsheet the company controls. Software becomes useful once several jobs are mapped or AI tools need to read the map automatically, and then a database, a small web app and version history are enough.

Map of Work: decide where AI fits in your business

Before you buy AI, write down how your experienced people actually decide, including the exceptions. A Map of Work is that record, owned by your company and usable by any AI tool.

Last updated 2026-10-11

TL;DR AI tools stall in small businesses because nobody wrote down how the work actually gets decided: the checks your best people make, the exceptions they allow, the mistakes they have learned to avoid. A Map of Work is that record, kept in a document your company owns. Start with one job, walk through a real case with the person who does it best, and write down what they check and why. Any AI tool you adopt later can read it, and the record stays with you when you change tools.

Most owners who try AI hit the same wall. The team uses a new tool for a week and drifts back to the old way, or the business is paying for AI and nobody can point to a single thing it changed. Sometimes the tool is the problem. It was built for a different kind of business, so the team bends the work to fit the software. Or the business runs five tools strung together with exports, copy and paste, and a spreadsheet in the middle, and nobody owns the gaps where things fall through. Adding one more AI tool to that stack adds one more gap.

Even the right tool, on its own, fails for a quieter reason: it was handed your documents, and your documents don't say how the work is done. Writing that down first is also what tells you which tool fits and how many you actually need.

The knowledge that decides outcomes lives in people's heads: which customer always pays late, which job looks simple and never is, when to ignore the price list. It also walks out the door. The median US worker had been with their current employer for 4.1 years in January 2026, according to the Bureau of Labor Statistics.

What is a Map of Work?

A Map of Work is a written record of how one job in your business actually gets done: the steps, the checks an experienced person makes, the exceptions they allow, why, and who said so. It belongs to the company, not to an AI vendor or a software platform, and people and AI tools can both read it.

The philosopher Michael Polanyi put the difficulty in one line in The Tacit Dimension (1966): "we can know more than we can tell." Your lead technician, your office manager or your best salesperson does dozens of small things that never made it into a procedure, and they would struggle to list them if you asked. A Map of Work gets them out one job at a time, by walking through real cases instead of asking for a brain dump.

Do not try to map the company. Pick one job that repeats every week and where experience visibly changes the result. Map that, use it, and only then pick the next one.

What does a Map of Work look like in a small business?

In a 25-person heating and cooling company, a single entry might record that the lead technician quotes a replacement option alongside the repair on any system over 15 years old that runs on R-22 refrigerant, why she does it, the exceptions, and how many callbacks the rule has prevented. The example below is illustrative.

The company's price book says a failed capacitor on a home air conditioner is a standard repair at a standard price. Junior technicians quote exactly that. Maria, the lead technician, checks two things first: how old the system is and which refrigerant it uses. US production and import of R-22 ended in January 2020, so recharging an old R-22 system is expensive, and those systems tend to fail again before the summer is out. When she sees one, she quotes the repair and a replacement side by side and tells the homeowner why.

The difference shows up in July, when the homeowner who got the cheap repair calls back angry and the one Maria quoted already has a new system. Written down, her habit looks like this:

Job Quoting a residential AC repair

The price book says Failed capacitor: standard repair, standard price.

What experience adds If the system is over 15 years old or runs on R-22, quote the repair and a replacement option side by side.

Why Old R-22 systems often fail again within the season, and the refrigerant alone can cost more than the repair.

Exceptions A replacement is already scheduled, or the customer is a landlord account with its own repair policy.

Source Maria, lead technician, walking through a June job.

Evidence Six callbacks last summer on old R-22 units that were quoted repair-only.

Approved The owner, July. Review again next spring.

Nothing in that entry needs software. It fits on an index card. But it now exists outside Maria's head: a new hire can learn it in their first week, the office can check quotes against it, and an AI tool that drafts quotes from a technician's notes can follow it. The tool would flag the old R-22 unit and attach the replacement option, and a person would still approve the quote.

The same shape fits most owner-operated businesses: a clinic deciding which appointment requests need a same-day slot, a distributor deciding when to waive a minimum order, a restaurant group setting prep quantities for an event weekend, a shop handling a return outside policy.

Why write it down before you buy AI?

Write it down first because an AI tool can only follow judgment that someone has recorded, and because a tool that watches your team work will otherwise hold the only written record of how your business decides. Keeping the record in a document you control lowers your dependence on any one vendor.

The first reason is practical. Tools that answer questions from your documents give generic answers, because generic is what the documents contain. Tools that take part in the work see a lot more: the goal, the information gathered, the decision, the human correction, the exception and the outcome.

The second reason is ownership. If the only place your know-how is written down is inside a vendor's platform, leaving that vendor means leaving the record behind. Disclosure is a related risk. After engineers pasted sensitive internal data into ChatGPT in April 2023, Samsung restricted generative AI tools on company devices, citing the difficulty of retrieving and deleting data held on external servers (TechCrunch).

A Map of Work does not make your information secret. Anything you send to an outside AI service has been sent. What it changes is that the record lives with you: a new tool can be pointed at the same document, and each tool can be given only the part a task needs.

Switching tools still costs something. You set up the new one, test it on real cases and fix what it gets wrong. But you don't have to rebuild what your company knows.

How do you build a Map of Work?

Build it one job at a time: choose a job where experience changes the result, gather what already exists, walk through a recent real case with the person who does it best, record every correction they make to someone else's work, and note where each rule came from. A shared document or spreadsheet is enough to start.

1. Choose one job

A useful test: if your three best people disappeared tomorrow, what would the business suddenly not know how to do? Among those jobs, start with the one that happens most often.

2. Gather what already exists

Price books, checklists, old quotes, emails, training notes, job tickets. Read them, then assume the real process is different. The checklist says the site survey always comes first; the last three jobs show that one skipped it and one did it after the quote. Those gaps are where the judgment lives.

3. Walk through a real case

Do not ask "tell me everything you know." Pick a recent job and go through it step by step with the person who handled it, the way the Critical Decision Method that Gary Klein and colleagues developed for studying expert firefighters walks through one real incident. Ask:

What did you look at first, and why?
What did you notice that told you this job was different?
What did you consider and decide against?
What would someone in their first month have done?
When would you do it differently?
Where did you learn that?

4. Record corrections

The moment an experienced person says "no, check this first" is the most valuable thing to capture, and it almost never reaches a document. Write down the situation, what the less experienced person proposed, what the expert changed, why, and how it turned out.

5. Note where each rule came from

Every entry says who it came from, what cases support it, who approved it and when to review it. This matters more once AI is involved, because AI makes it easy to produce rules that sound authoritative and came from nowhere. An AI-generated inference should never silently become company policy; a person approves it first.

Who keeps a Map of Work current, and what does it cost?

Each map needs one person responsible for it, usually the manager of that job, and it is updated whenever a correction happens and reviewed on a schedule, for example every season. The cost is mostly the expert's time: one walkthrough to start, then a few minutes per correction.

A map nobody updates becomes another stale procedure. The cheapest habit that keeps it alive is to add an entry, or change one, every time an experienced person overrides the usual way. Over a season those entries show which rules hold up and which ones keep needing exceptions. An exception that happens 40% of the time is not an exception anymore, and the rule should change.

Not every job needs a map. If no one's judgment changes the result, a plain checklist is enough. Save the effort for the jobs where your best person and your newest person get different outcomes.

How is a Map of Work different from an AI knowledge base?

An AI knowledge base loads your documents so a tool can answer questions about them. A Map of Work adds what the documents leave out: the judgment, exceptions, corrections and outcomes behind how your people decide.

A knowledge base can tell a new employee what the company knows. A Map of Work begins to teach them how the company decides.

What can you use a Map of Work for once it exists?

The same record serves training, quality checks, succession planning and any AI tool you adopt, because none of them owns it. Each one reads the same approved entries, so improving the map improves all of them.

Training. A new hire reads the entries for their job and the cases behind them, instead of shadowing someone for months.
Quality checks. The office compares quotes, intakes or orders against the map and catches the ones that skipped a check.
Succession. The map shows which jobs depend on one person. The Pew Research Center estimated that about 10,000 US Baby Boomers would turn 65 every day for 19 years starting in 2011, so in many small businesses the person who knows the most is close to retiring.
AI tools. A tool that drafts, sorts or checks work can be told which rules apply and when to hand a case to a person.
Decisions about automation. A job whose map has stopped changing is a candidate for automation. A job that still collects exceptions every week needs a person.

Thinking about where AI fits in your business? Send me one job your team does every week where experience changes the result. Bring the messy version: how the job really runs, not what the procedure manual says. You get a one-page answer by email: what I would map first, and whether AI belongs in that job at all. There is no charge and no call. If you find yourself coming back with more questions like it, that is what an engagement is for. Send me the job

Rather work through it together? A Working Session is 90 minutes on that job and anything else on your mind, with written notes afterwards, for $750.

Worksheet: your first Map of Work

Use this one-page worksheet to map your first job. Print it, take it to a walkthrough with the person who does the job best, and fill one copy per rule.

My first Map of Workbigballi.com/map-of-work

The job. Which job, if our best people left tomorrow, would we no longer know how to do?
The case. A recent, real example of that job: who handled it, when.
The official way. What the price book, checklist or procedure says to do.
What experience adds. What the expert checked or did differently.
Why. What goes wrong when it is skipped.
Exceptions. When the expert would not do it.
Evidence. Past jobs that show it works, or callbacks that show what happens without it.
Source and approval. Who it came from, who approved it, when to review it.

For the technical reader: how the Map is built

Whoever builds the system around the Map can store entries as structured text, record decisions as they happen, keep the Map separate from any AI model, and send each model only the context one step needs. None of this needs an enterprise platform.

Show the technical detail

Entries as structured text

An entry like Maria's can live as YAML or JSON, which people can read and any tool can parse. Plain structured text keeps the record yours: it moves between databases, tools and model vendors with little more than reformatting.

rule:
  id: R12
  job: quote_residential_ac_repair
  situation:
    - system_age_years > 15
    - refrigerant: R-22
  action:
    Quote the repair and a replacement option side by side.
  rationale:
    Old R-22 systems often fail again within the season;
    recharge cost can exceed the repair.
  exceptions:
    - replacement_already_scheduled
    - landlord_account_with_own_policy
  source: lead_technician
  evidence: [job_4412, job_4420, job_4517]
  approved_by: owner
  review: next_spring
  status: approved

Decisions recorded as they happen

Each consequential decision can be logged with the case, what the tool or junior person proposed, what the person in charge decided, the rule applied and the outcome. Over a season those records show how often each rule holds and where its exceptions cluster.

decision:
  case: job_4630
  proposed: repair_only
  decided: repair_plus_replacement_option
  decided_by: lead_technician
  rule: R12
  outcome: replacement_booked

Separate the Map from the models

The company's system orchestrates the work and sends a model single operations such as extract, classify or summarize, with only the context that step needs. This limits what each call discloses. It does not prevent disclosure: a provider that handles many calls over time can still learn a great deal, and local models reduce that exposure only partly (see whether local models protect your know-how). Open standards such as the Model Context Protocol, released by Anthropic in November 2024, make it practical to connect one context layer to different models.

Provenance and audit

Each entry and decision keeps its source, approver and version, the idea the W3C PROV standard formalizes. That makes a decision reconstructable after the fact, which the NIST AI Risk Management Framework treats as part of governing AI systems.

A minimal stack

A first version runs on a relational database such as PostgreSQL or SQLite, file storage for source documents, a small web app for interviews and approvals, version history, and a thin layer that lets you swap model providers. Its jobs are to ingest existing material, support the walkthrough interviews, structure entries, route them for human approval, capture new decisions and export the Map for people, agents and automation.

Key takeaways

AI tools stall when they only see your documents; the judgment that runs your business is rarely written down.
A Map of Work records how one job is actually done: what experienced people check, the exceptions, why, and who said so.
Start with one job, walk through a real case with your best person, and keep the record in a document you own.
Owning the record lowers your dependence on any AI vendor. It does not make information you send to that vendor private.

Frequently asked questions

What is a Map of Work?: A Map of Work is a written record of how one job in a business actually gets done: the steps, the checks an experienced person makes, the exceptions, the reasons and the source of each rule. The company owns it, and both people and AI tools can use it.
How is a Map of Work different from an AI knowledge base?: A knowledge base loads documents so a tool can answer questions about them. A Map of Work also records judgment: why an expert rejected the obvious answer, which exceptions apply and how past decisions turned out.
Where should a small business start?: Start with one job that repeats every week and where experience changes the result, such as quoting a repair, triaging appointment requests or approving an exception to a policy. Walk through a recent real case with the person who does it best and write down what they check and why.
Does a Map of Work stop AI vendors from seeing company data?: No. Anything sent to an outside AI service has been disclosed to that provider. A Map of Work reduces dependence on one vendor and lets each tool receive only the context a task needs, but over many requests a provider can still learn a lot about how you work.
Do you need special software to build a Map of Work?: No. A first map fits in a shared document or spreadsheet the company controls. Software becomes useful once several jobs are mapped or AI tools need to read the map automatically, and then a database, a small web app and version history are enough.

Related guides

Sources

Michael Polanyi, The Tacit Dimension (1966).
Ikujiro Nonaka, "The Knowledge-Creating Company" (Harvard Business Review)
Gary Klein and colleagues, the Critical Decision Method for eliciting expert knowledge.
Bureau of Labor Statistics, Employee Tenure
Pew Research Center, Baby Boomers Retire
TechCrunch, Samsung restricts generative AI tools
W3C PROV overview
Model Context Protocol
NIST AI Risk Management Framework

About the author

Giacomo Balli is an independent technology advisor in San Francisco who has built and shipped software since 2010. He helps owners decide where AI belongs in how their company works, and what the company should own before a vendor or platform learns it.

Disclosure

Disclosure: Giacomo Balli is paid only by the owners he advises, through a fixed-fee advisory engagement or a one-off Working Session. He does not resell AI tools or development work, take equity, or earn vendor commissions. Company, product and personal names in the examples are illustrative, not endorsements or real clients.