Source: [AI won’t fix a mess. It makes more mess.](https://treenodes.com/notes/ai-will-not-fix-a-messy-process/) This Markdown version is generated from the public page. Visit the source for interactive examples and full-size visuals. [Notes](https://treenodes.com/notes/) Before you add AI # AI won’t fix a mess. It makes more mess. **Fix the process. Then put AI to work.** Point AI at an unclear process and it won’t stop to ask. It will automate your mistakes, at volume. [Nermien Barakat](https://treenodes.com/articles/by/nermien-barakat/), Principal Software Architect & Engineer · 2 October 2026 · 5 min read Illustrative example: one call-off at a supplier to trade and contractor customers, working to order or from stock. [Read the note](https://treenodes.com/notes/ai-will-not-fix-a-messy-process/#note-text) From the problem to the fix ![Headline: AI won’t fix a mess. It makes more mess. Below: Fix the process. Then put AI to work. Label: uncontrolled environment. A loose black line labelled by hand, unclear steps, a person can stop and ask, runs into a purple AI circle and leaves as a red coil of five loops labelled with AI or software, but no architecture, testing or automation: more mess, faster. Footer: AI does not ask unless told when to. Where the process is unclear, it guesses.](https://treenodes.com/assets/notes/slides/ai-will-not-fix-a-messy-process/slide-1.webp) **AI won’t fix a mess. It makes more mess.** Unclear steps slow a person down. Given to AI with no rule for when to ask, they multiply. ![Illustrative. Four boxes in a row, order, production or stock, delivery plan and invoice, each holding the same quantity of 20 typed again. Notes: the date moved twice by email, the plan still showing the first date, the customer asking where it is, accounts asking for the signed note.](https://treenodes.com/assets/notes/slides/ai-will-not-fix-a-messy-process/slide-2.webp) **One quantity, typed four times** Four copies of one number, a date moved twice and two people chasing. Nobody has added AI yet. ![Three cards naming the causes: no rule for which price wins, nobody owns the delivery date, and no check before confirming. Footer: AI can sort messy information. It can’t settle a messy process.](https://treenodes.com/assets/notes/slides/ai-will-not-fix-a-messy-process/slide-3.webp) **Why it stays slow** None of these is a reading problem. Each is a decision nobody has made yet. ![Illustrative. Call-off A, same again Tuesday morning: 12 called off, read as a repeat of the 9 September call-off, 12 ready, items, price and slot checked against named records, passes to one-click approval. Call-off B, 20 for Thursday: 20 called off, 15 ready, 5 short, stops; the operations manager rings the customer. Footer: Thursday is requested; delivery is not yet agreed.](https://treenodes.com/assets/notes/slides/ai-will-not-fix-a-messy-process/slide-4.webp) **One passes. One stops.** A matches every record and waits for one-click approval. B is five short, so a named person decides. ![Two labelled lanes. Stays inside the business and can be undone, sampled afterwards: sorting incoming messages, matching delivery notes to orders. Leaves the business or can’t be undone, a named person or deputy first: confirming a delivery date, releasing work, sending an invoice or application, messages about price, delay or scope. Footer: You decide which lane each step sits in, and the software enforces it.](https://treenodes.com/assets/notes/slides/ai-will-not-fix-a-messy-process/slide-5.webp) **Match the review to the risk** Sorting can be checked afterwards. Anything that leaves the business or can’t be undone waits for a named person. ![Heading: Nothing goes live on a promise. It goes live on a measured result. Six numbered steps: measure today; set the pass mark first, strict on a missed difference; build an answer key from a closed job under today’s rules; keep some cases unseen; run alongside the team with nothing sent; go live with approval, sampling and a switch back. Last line: Start with one call-off type.](https://treenodes.com/assets/notes/slides/ai-will-not-fix-a-messy-process/slide-6.webp) **Nothing goes live on a promise** Strict on a missed difference, tolerant of the odd false alarm. Nothing is sent or released during the trial. The note in 1 minute Video · no sound The six slides as a silent one-minute walkthrough of the call-off example. The written note explains each step. Michael Hammer’s 1990 warning still holds: don’t automate a bad process, fix it. AI makes it sharper. Where a step is unclear, a person stops and asks. A model works on probability: unless the task says when to stop, it gives the most likely answer, confidently and at volume. Take an ordinary week. A customer says yes. The same quantity is typed four times: into the order, production or stock, the delivery plan and the invoice. The date moves twice by email. The plan still shows the first. The customer chases the delivery. Accounts chase the signed note. Now add AI. A repeat order arrives at last quarter’s price. AI reads it perfectly. Nobody has decided which price wins, so the system confirms the old one, faster. The observation ## Why it stays slow A common pattern: the delay is rarely in reading the order. It sits in handoffs and in three open decisions: which price wins, who owns the delivery date, and what is checked before confirming. If nobody has agreed what happens when a delivery is short, a faster system reaches the same unresolved decision sooner. AI can sort messy information. It can’t settle a messy process. Proposed solution ## One call-off, done properly ### Where AI earns its place - Reading a call-off into one draft, each field linked to its source, whether it arrives as an email, a spreadsheet, a photo of a handwritten list or “same again, but 20 for Thursday”. - Sorting what arrives into enquiries, new call-offs, amendments and chasers, and proposing which order each belongs to. Some of this needs no AI. Entering a quantity once and listing every unanswered message is basic record-keeping. AI works inside ordinary rules, run by software, that decide when work passes as a draft and when it stops. Here, a confirmation is drafted only if the price matches the agreed order, the quantity is ready for the date and the slot is free. Anything else stops for a named person. These checks are the definition of done, because AI works on probability: its answer is likely, not certain, even when it sounds sure. A model can suggest a rule; your business decides which rule applies. ### Who does what 1. Call-off in **AI proposes** Sorts it, proposes which order **Software checks** Checks sender and order, logs it once **Records prove** Original message, agreed order **People commit** Sales office, if unmatched 2. Checked before confirming **AI proposes** Drafts the confirmation from the checked figures **Software checks** Checks price, quantity ready and slot **Records prove** Agreed order, production record, delivery plan **People commit** Operations manager approves every confirmation 3. Delivered and invoiced **AI proposes** Reads the signed note, handwriting included **Software checks** Checks it against dispatch, drafts the invoice line **Records prove** Signed delivery note, dispatch record **People commit** Accounts approve; operations if signed short ### One passes. One stops. Illustrative Call-off A “Same again, Tuesday morning.” - Items and price match agreed order - Tuesday morning slot free delivery plan **Called off** **12** customer email, read as a repeat of the 9 September order **Ready** **12** production record **Passes:** drafted for approval. The operations manager approves with one click. Call-off B “20 for Thursday.” **Called off** **20** customer email **Ready** **15** production record 5 short **Stops:** the operations manager rings to offer 15 on Thursday or all 20 later. AI drafts the follow-up; the manager sends it. Nothing was misread. The check against the production record stopped B, not any doubt in the model. Thursday is requested; delivery is not yet agreed. Show the approver the evidence, not just a button. Passing checks isn’t proof. Suppose A’s customer meant their second site, and AI linked “same again” to the first. Every check passes, because each tests the order AI picked. Sampling what passed finds it. The fix is a rule: a repeat from a customer with two open orders stops. ### Match the review to the risk Stays inside the business, can be undone Sampled afterwards - Sorting incoming messages - Matching delivery notes to orders Leaves the business or can’t be undone A named person or deputy approves first, even an exact repeat - Confirming a delivery date - Releasing work to production - Sending an invoice or application - Any message about price, delay or scope You decide which lane each step sits in, and the software enforces it. Nothing that promises a date is sent automatically. What arrives is evidence for a person, never a command for the system. A customer’s “same price as last time” or a site manager’s “just do it” is logged for someone to confirm in writing, price or query. It can start a draft, never skip a check or approve anything. ### The same pattern for contractors and fabricators Illustrative A site engineer emails a rebar fabricator: steel for pour 3 now wanted Tuesday, Revision D schedule attached. AI reads the email into a call-off. Software checks the register shows D as current, then compares it with what is already cut. Nothing more is cut or loaded. **Already cut to Revision C** **14.2 t** production record **Needed for Revision D** **15.1 t** six bar marks changed · bar schedule, drawing register Its contracts manager decides what to cut and agrees a date with site; its QS decides whether to price the re-cut as a variation. Design questions go to the designer as an RFI, through the main contractor. The test ## How you’d know it works ### A trial with a pass mark 1. Measure today: confirmations that didn’t match the order, deliveries rebooked or turned away, disputed quantities and credit notes. 2. Set the pass mark first: strict on a missed difference, tolerant of the odd false alarm. 3. Build an answer key from a closed job, reviewed under today’s rules, not copied from what was sent. 4. Keep some cases unseen until the final score. 5. Run it alongside the team, nothing sent or released. 6. Go live on one call-off type, with one-click approval, sampling and a switch back. Retest when the model or rules change. The next step ## Where to start Follow one recent order from the customer’s yes to the signed delivery note. Count how often the quantity was typed, who touched or chased it, where it waited, and which decision had no named owner. ### Three questions 1. Can you say within a minute what each customer was promised, and at which price? 2. Is every change made by phone or on site written down, with a name against it? 3. When a delivery will be short, is it clear who decides what the customer is told? Start with the first no. Where does your paperwork fall furthest behind the work: the order, the delivery or the changes agreed along the way? Bring one request that keeps getting passed around. We’ll map the records, decisions and handoffs it needs. [Let’s map it](https://treenodes.com/contact/) Go deeper [Giving AI a job description](https://treenodes.com/articles/giving-ai-a-job-description/) Before AI reads an invoice or drafts a reply, decide what it should do when it gets stuck. Define the task, the review and the fallback. Sources 1. [Reengineering Work: Don’t Automate, Obliterate (opens in a new tab)](https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate) Michael Hammer, Harvard Business Review, July–August 1990 2. [Humans and Automation: Use, Misuse, Disuse, Abuse (opens in a new tab)](https://doi.org/10.1518/001872097778543886) Raja Parasuraman and Victor Riley, Human Factors, vol. 39, no. 2, 1997 3. [Hidden Technical Debt in Machine Learning Systems (opens in a new tab)](https://papers.nips.cc/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html) D. Sculley et al., Advances in Neural Information Processing Systems 28, 2015 4. [Building effective agents (opens in a new tab)](https://www.anthropic.com/engineering/building-effective-agents) Anthropic Engineering · workflows, tools and evaluation