The Sixteen Intelligence Units Behind Business Processes
Every business process is a chain of small judgements. We count sixteen kinds, and this is what each is responsible for, which need an AI model and which do not, and how far they go.
When a company asks us to "add AI" to a process, the first thing we do is stop treating the process as one job. An invoice arriving by email, a customer asking for a price, a complaint about a late delivery: each is a chain of small, separate judgements. Someone reads a document. Someone works out which supplier it belongs to. Someone checks it against the order. Someone decides whether it can be paid.
We call each kind of judgement an intelligence unit. We count sixteen of them. This article sets out what each one is responsible for, which of them need an AI model and which do not, and how far the sixteen go towards covering the work of a business.
One rule decides who does the work
Every unit can be done in one of three ways: by an AI model, by ordinary code, or by a person. We choose with one rule.
- Use a model where the input is messy or the judgement is fuzzy: a scanned invoice, a rambling email, an item described in a supplier's own words.
- Use code wherever the rule can be written down: a credit limit, a duplicate check, a reorder point.
- Keep a person where the decision carries accountability: sending the quotation, posting the invoice, shortlisting the candidate.
This is why four of the sixteen units use no model at all. A language model can be persuaded, and it is poor at arithmetic over time. A rule in code gives the same answer to the same facts every time.
The sixteen units
The numbers are the identifiers we use in our own designs. Select a unit to see what it is responsible for, with two examples from different kinds of business. The examples are illustrations of where the unit applies, not results from client projects.
AI modelCode and people, no modelSelect a unit to see its job and two examples.
Reading what arrives
2 · ExtractionSmall to mid model
Turns a document into structured fields: the supplier, the date, each line and amount on an invoice, a purchase order or a CV. The output follows a fixed schema, and the ERP validates every field before it is saved.
For example. Trading: a supplier invoice arrives as a PDF, and the supplier, date, each line and the totals land in a draft purchase invoice. Recruitment: a CV becomes a candidate record with employers, dates, qualifications and skills in separate fields.
3 · Classification and routingSmall model
Decides what something is and where it goes: which inbound emails are orders, which are complaints, what language they are in and who should pick them up. It is usually the cheapest unit.
For example. Distribution: a shared sales inbox is sorted into orders, quotation requests and complaints, each sent to the right queue. Property management: tenant messages are sorted into maintenance, billing and renewals, with urgent ones marked.
14 · Visual understandingMid model with vision
Reads images and drawings: a photo of a damaged delivery, a receipt, a picture of a part that has to be matched to the catalogue.
For example. Logistics: a photo of a crushed carton at goods receipt is recorded as damage on the receipt, with the type of damage. Automotive service: a photo of a worn part is matched to the part in the catalogue.
15 · TranslationSmall to mid model
Works across languages and vocabularies. That includes a supplier's jargon against your internal terms, not only Arabic against English.
For example. Contracting: Arabic tender documents are put into English for the estimating team. Import trading: a Chinese supplier's product names are mapped to the terms your own team uses.
Connecting it to your records
1 · MatchingMid model
Links an outside description to the right internal record: a supplier's item description to your item code, a bank line to an invoice, two customer records that are the same company. Code shortlists the candidates. The model picks one and gives its reason.
For example. Building materials: "OPC cement 50 kg bag" on a supplier quote is linked to your own item code. Finance: a line on the bank statement is linked to the invoice it pays, even when the reference is incomplete.
4 · Normalisation and enrichmentSmall model
Cleans and completes records to a standard: units of measure, item names, addresses, missing attributes. Code handles units and formats. A model handles free text.
For example. Retail: product names, sizes and colours from many suppliers are brought to one format before they go on the website. Manufacturing: "pcs", "each" and "nos" become one unit of measure, and missing material grades are filled in from the description.
5 · ValidationMostly code, small model
Spots what is wrong or unusual: a duplicate invoice, a price that differs from the order, an odd expense claim. Thresholds and duplicate checks are code. A model sees only the cases code cannot settle.
For example. Finance: the same invoice sent twice under two different numbers is caught before it is paid twice. Field services: an expense claim for fuel on a day with no site visit is flagged for a second look.
10 · Knowledge retrievalMid model
Answers from company records and documents, and cites the source: what you quoted this client last year, what the policy says about a return. Search finds the passages. The model writes the answer.
For example. Professional services: "What did we charge this client for last year's audit?" is answered from past quotations, with a link to the document. HR: a staff member asks how much leave carries over, and the answer cites the clause in the policy.
Producing and communicating
11 · DraftingMid to top model
Produces a first version for review: a quotation, a reply to a customer, a job description, a report. A person reads it before anything is sent.
For example. Engineering supply: a quotation is drafted from the customer's enquiry and the current price list. Retail customer service: a reply to a late-delivery complaint is drafted with the order's real status in it.
13 · Conversational intakeMid model
Gathers missing information from people by talking to them: asking a customer what they need, or chasing a supplier for a missing specification. The conversation collects the facts. It does not decide what happens next.
For example. Car service centre: a WhatsApp conversation collects the model, mileage, the problem and a preferred time before the booking is made. Procurement: a supplier is asked for the lead time and test certificate missing from their quote.
12 · Summarising and monitoringSmall to mid model
Condenses activity and flags what needs attention: meeting notes into actions, a daily digest of exceptions. Code detects the events. A model writes the digest.
For example. Consulting: a recorded client meeting becomes a list of actions, each with an owner. Distribution: a morning digest lists late deliveries and items that have fallen below their reorder level.
6 · Comparison and evaluationTop model
Weighs options against criteria: three supplier quotes, a contract against your standard terms, a candidate against a role. This is the one unit where we reach for the most capable model. It drafts the comparison, and a person decides.
For example. Contracting: three subcontractor quotes are set side by side on price, delivery and exclusions. Legal review: a supplier's contract is compared with your standard terms, and the clauses that differ are listed.
Deciding and carrying through
7 · Policy decisioningCode and people
Applies rules to approve, reject or escalate: credit limits, discount approvals, expense claims, whether a lead qualifies. The rules are written or configured. A model may assemble the facts and the relevant policy clause, but it does not decide.
For example. Wholesale: an order is released if the customer is inside their credit limit and held if not. Any employer: an expense claim under the limit is approved, and one above it goes to the manager.
8 · ForecastingStatistical code
Predicts a number or a date: demand, cash flow, a delivery date. Language models are poor at numeric forecasting, so this is statistics. A model can explain the result in plain words.
For example. Food distribution: next month's demand for each item is predicted from sales history and season. Finance: the cash position for the next eight weeks is predicted from invoices due and how each customer usually pays.
9 · Planning and schedulingSolver code and people
Works out a feasible sequence under constraints: reordering, production schedules, shift rosters, delivery routes.
For example. Manufacturing: jobs are sequenced around machine capacity and promised dates. Field service: the day's technician visits are ordered into routes that fit the booked time slots.
16 · Follow-throughWorkflow code
Carries a multi-step task across systems to completion: reminders, retries, hand-offs, escalation when something is stuck. This is orchestration, in a workflow tool such as n8n, and it is the glue that calls the other units.
For example. Finance: an unpaid invoice gets reminders at set intervals, then goes to the account manager. HR: a new joiner's tasks across IT, HR and payroll are tracked until every one is done.
A process is a chain of units
Once the units have names, a process can be written as a short chain. Two examples from our own designs:
- A supplier invoice arrives by email. Extraction reads it. Matching ties each line to an item and the invoice to its order. Validation checks for duplicates and price differences. Policy decisioning applies the approval rule. A person posts it. Follow-through chases anything outstanding. A model is involved in the first two steps and part of the third.
- A customer asks for a quotation. Classification recognises the enquiry. Conversational intake asks for what is missing. Your qualification rules run in code. Matching maps the requested items to the catalogue. Drafting prepares the quotation. A person reviews and sends it.
Writing a process this way shows where the cost and the risk sit. It also shows how little of the chain is "AI" in the sense people expect. Most steps are small, several are plain code, and the decision stays with a named person.
How far sixteen units go
Our working estimate is that these sixteen units cover about nine in ten of the steps in the routine processes of a business. That is a judgement from the processes we have mapped. It is not a measured figure, and we would rather say so than dress it up.
What we can show is narrower and concrete. In our automation cost estimator, twenty workflows across sales, finance, customer service, HR, procurement and operations are each built from these units, a person's review, and a few supporting steps: reading scanned pages, transcribing speech, screening what goes in and out, and, in one workflow, an agent that calls tools. You can open any workflow there, change its steps or build your own from the same units.
The units cover so much because routine office work repeats a few moves: read something, connect it to a record, check it, decide by a rule, write something, pass it on. The department changes. The moves do not.
What the units do not cover
- Accountability. A unit can prepare a decision. A person still answers for it.
- Relationships and negotiation. Winning a client's trust or settling a dispute is not a chain of small judgements.
- The new and the one-off. A situation with no precedent has no rule to code and no examples to test a model against.
- Physical work. Units can plan a delivery. They cannot make it.
From a unit to an agent
An agent, as we build it, is one unit given a model, rules and limits. Every agent is specified with the same eight fields.
- Model: the model and size assigned.
- Rules: what it must do, step by step where needed.
- Constraints: what it must never do.
- Character: tone and manner, where it talks to people.
- Contract: the exact input and output, so units can be chained and swapped.
- Tools: what it may read and write. It may read the item master, for example, but never submit a document.
- Confidence and escalation: the point below which it hands the case to a person, and to whom.
- Test set: thirty to fifty real examples with known answers.
The test set matters most. It turns "a suitable model" into a measurement, and it is what later allows a move to a smaller, cheaper model with evidence that accuracy holds.
Matching shows a second effect. Every match a person confirms is saved as an alias, so the same supplier description is recognised next time without calling a model. The cost of the unit falls as the business uses it.
Why this is the right size of piece
- It can be priced. Each unit has a volume, a typical size and a model tier, so the cost of a process is the sum of its units. That is how the estimator works.
- It can be tested. A unit has one job and a known right answer, which a whole process does not.
- It can be replaced. A better or cheaper model can take over one unit without touching the others.
- It keeps models where they belong. Rules stay in code and decisions stay with people, by design and not by hope.
Where to start
Take one process and write it as a chain. Mark each step as model, code or person. The steps marked "model" are your first candidates, and the smallest of them with a clear right answer is the place to begin.
To see what a chain would cost to run, try the automation cost estimator. To talk through one of your own processes, get in touch.
Vivek Digambernath
Harmonizing the past, present, and future through music, technology, and travel.
From the article to a working brief
No comments yet. Login to start a new discussion Start a new discussion