What it is
Before you set up an AI tool that learns from your own history, check the records it will learn from. The records need to be complete, recorded the same way each time, and held in one place. If they are not, fix that first.
Such tools estimate the price of a new job from past jobs, the time a task will take from past tasks, or the next step with a customer from that customer’s record.
A general chat assistant used for drafting does not need this step. A tool that predicts from your own data does.
What the evidence shows
The company cases listed under “Company cases” below are the basis for this advice. No study has tested it.
- The clean-up was the main work. A six-person metalworking shop wanted AI to price custom parts from its production history. A large part of the project went on preparing the records. The owner says: “If the data is not clean, the tool will not be reliable.” A quote now takes about ten minutes where it sometimes took two hours.
- The records had to be created. A steel fabricator had production times written down by hand. The times for identical parts differed by up to 30 minutes, so they were of little use. The first half of the project was a simple app for workers to log real times. The AI model trained on those times was about five times closer to the real figure than the firm’s experts.
- One place for each customer. A two-person advisory firm ran on emails and spreadsheets. The firm names one file per client, in place of scattered files, as one of the choices its project depended on. A cleaning company kept its customers in spreadsheets. The company moved its customers into one customer system, and the AI in that system now suggests replies from each customer’s record.
- Poor records in, wrong answers out. Rite Aid let low-quality images into the database its facial recognition system matched shoppers against. The regulator said this made false matches more likely. Shoppers were wrongly accused, and the chain was banned from the technology for five years.
Where long, complete records already exist, this step is short. A chip plant had the records of 1,197 past production batches, each with 3,614 recorded values. An AI model trained on them pointed to two production steps that were spoiling chips.
An AI vendor makes a similar point from its own data. Anthropic, which sells the AI model Claude, classified how businesses use the model. In its September 2025 report it argues that complex use is held back more by access to the right context than by the model itself. That is the vendor’s reading of its own customers’ use, not a test.
How to do it
- Name the one thing you want predicted. The metalworking shop wanted the price of a new part. The steel fabricator wanted the time to roll one sheet. One prediction tells you which records matter.
- List the records the prediction needs and where they are. For the metalworking shop these were the entries in its business software, the parts lists, the 3D files and the standard manufacturing routes.
- Test whether the records can be trusted. Take ten identical or near-identical past jobs and compare what was recorded. The steel fabricator found gaps of up to 30 minutes between identical parts. That told the fabricator to start again.
- If the record is not good enough, start recording. Keep it simple. The steel fabricator broke one process into steps and gave workers an app to log each step against the job. The fabricator adjusted the app with the workers and trained them on it.
- Record what really happened. The fabricator’s experts had been overestimating side tasks such as moving material. A tool trained on estimates would have copied that error. Train on measured times, real costs and real outcomes.
- Link each record to its job or customer. The metalworking shop linked each 3D file to the record of how the part was made. The advisory firm put everything about a client in one file.
- Decide what may go in. Set a minimum standard for a new record and name who checks it.
- Keep recording after the tool is live. The steel fabricator still logs every new job and updates its model from those records at regular intervals.
What to measure
- The gap between records of identical jobs, before and after the clean-up.
- The share of past jobs with a complete record.
- How far the tool’s predictions are from what really happened, on jobs it has not seen. See measure against a baseline.
Where it fails
- Cleaning everything. The steel fabricator limited its project to one process. Clean the records one prediction needs, then move on. See start with one costly process.
- Recording that staff do not keep up. The centre that ran the steel fabricator’s project reports that people problems kept slowing the data collection and stretched the project out. Make logging quick, and involve the people who do it.
- Nobody owns the records afterwards. New entries slip back to the old standard. Name one person.
- Starting with the tool. A tool bought before the records are checked will be judged on answers it could not get right.
By company size
- Micro and small: the owner can do this. The metalworking shop’s owner ran the project himself with his assistant and no consulting firm.
- Medium: pick one process and give its records to one named person. Expect collecting and cleaning to take a large part of the project.
- Mid-market and enterprise: the records usually exist, spread over several systems. The work is linking them and agreeing one way of recording across sites and teams.
Company cases
Companies that did this
- EIST Before the AI was used, the firm cleaned the records in its business software, structured its parts lists and linked each 3D file to the record of how the part was made.
- Herges Stahl- und Blechbau The first step of the project was a simple app for workers to log real production times, because the hand-kept records differed by up to 30 minutes for identical parts.
- Ministry of Clean The company moved its customer list out of spreadsheets into one customer system, and the AI suggests replies and next steps from the customer data held there.
- SAM The firm names centralised data, one file per client in place of scattered files, as one of the choices that decided the project’s success.
What happened without it
Sources
No published study tests this practice directly, so it is rated Emerging. The rating rests on the company cases above and on the sources below.
- Anthropic (15 September 2025). Anthropic Economic Index report. A report by an AI vendor on how its own customers use its model. anthropic.com
Related practices
- Start with one costly process, not a company-wide rollout Pick one process that visibly costs time or money. Apply AI there, with a person checking the output, and expand only when the numbers move. Do first · Effort: medium (weeks, a few people)
- Measure time saved per task against a baseline Record how long a task takes and how good the result is before you add AI. Without that starting figure you cannot tell a real gain from a hopeful one. Do first · Effort: low (days, one person)
- Budget for people and process, not just tools Licences are the small part of the cost. Plan most of the time and money for training, redesigning work, and helping people through the change. Do first · Effort: low (days, one person)
- Decide what data may go into which AI tools Sort your data into two or three groups and say which AI tools each group may go into. Client secrets and personal data stay out of public tools. Where a mistake would be serious, block the upload instead of relying on a rule. Do first · Effort: low (days, one person)