Operational intelligence is usually described in the language of large organisations: data platforms, analytics teams, control towers. A business of ten or fifty people needs the same thing — information that turns into decisions, and decisions that turn into action — but not the same machinery. This guide sizes the operating layer described in Operational intelligence for a small team: a handful of decisions, the tools you already have, one page and one weekly meeting.
The problem: many tools, no time to look
A small business typically runs on several systems: accounting, a CRM or sales tracker, a shop or booking system, a support inbox, a few spreadsheets, perhaps a project tool. Each holds part of the picture. Nobody has time to look at all of them, and when someone does, the numbers disagree.
The result is decisions made late, or by feel. Stock runs out because nobody noticed it falling. A good customer goes quiet and nobody calls. An invoice sits unpaid for two months. The data to prevent each of these existed; nobody was looking at the right number at the right time.
Concept: the loop, sized small
Operational intelligence is a loop with five stages: signals from your systems, interpretation that turns data into meaning, decisions with clear owners, actions that carry them out, and measurement that shows whether it worked.
For a small team the loop does not need a platform. It needs:
- five decisions, not fifty metrics;
- one owner for each decision;
- one page that says what triggers each decision and what happens next;
- one weekly review where the loop is checked and improved.
Dashboards tell you what happened. The loop decides what to do about it.
When to use it
This guide is for you if:
- the weekly meeting starts with an argument about which number is right;
- customers notice problems before you do;
- the owner or one senior person is the only one who "just knows" what is going on;
- you have bought reporting tools that nobody opens;
- you are considering AI but are not sure what it would actually do.
Architecture: the minimum stack
- Your existing systems — accounting, sales, bookings or shop, support. These are the signal sources. Do not replace them.
- One place for the numbers — a shared spreadsheet or simple dashboard holding the handful of figures you need, refreshed automatically where possible.
- The decision register — one page listing each decision, its owner, its signal and what happens next.
- Alerts — a message to the owner when a condition is met, so nobody has to remember to look.
- The weekly review — thirty minutes, the same agenda every week.
Components
The decision register
The register is the heart of the system. Each row is one recurring decision.
| Decision | Owner | Signal | Condition | Default action | By when |
|---|---|---|---|---|---|
| Reorder stock | Operations lead | Stock level per product | Below four weeks of average sales | Raise a purchase order | Same day |
| Chase a payment | Finance lead | Invoice age | Unpaid 7 days after due date | Send reminder; call at 21 days | Within 2 days |
| Call a quiet customer | Account owner | Days since last order | Twice their usual interval | Personal call or email | Within a week |
| Escalate a complaint | Support lead | Ticket tags and age | Complaint unresolved after 48 hours | Owner contacts customer | Same day |
| Review pricing | Managing director | Margin per product line | Margin below target for a month | Review at next weekly meeting | Weekly |
Illustrative register, not a client engagement. Your decisions, thresholds and timings will differ. What matters is that each row is specific enough that anyone could tell whether it was followed.
The weekly review
The review turns the register from a document into a habit. Keep the agenda fixed and short:
- What fired? Each condition met this week, and who was told.
- What was done? The action taken, and whether it happened in time.
- What is still open? Anything waiting, and who owns it.
- What should change? Thresholds that fired too often or too late, signals nobody acted on, decisions that need a new owner.
If the review regularly runs past thirty minutes, the register holds too many decisions or the conditions are too loose.
Where an agent fits
An AI agent fits the watching and preparing parts of the loop:
- checking the signals continuously and telling the owner when a condition is met;
- assembling the context for a decision — the customer's history, open invoices, recent tickets — into a short brief;
- preparing the weekly summary: what fired, what was done, what is still open;
- drafting routine follow-ups for the owner to approve.
It should not make the register's decisions on its own at first. Pricing, credit, customer commitments and anything irreversible stay with a person. This is the role Oracle plays in the Guild: it watches operational data and surfaces what changed, why it matters and who should know, inside the Citadel and supervised by a Human Wizard.
Implementation
- Pick five decisions. The recurring ones that cost money when they are late — not the most strategic, the most frequent with a visible cost.
- Find the signal for each in a system you already have. If the data does not exist, start recording it or choose a different decision.
- Write the register and name the owners. Agree conditions and default actions with the people who will own them.
- Run the weekly review by hand for a month. For each decision: did the signal fire, did the owner act, did it work? Adjust thresholds that fire too often or too late.
- Then automate the reminders. Set up alerts so owners are told when a condition is met, and connect the numbers so nobody copies them by hand.
- Then consider an agent — reading data and recommending first, with its role widened only when the evidence shows it is reliable.
Failure modes
- Too many metrics. Fifty numbers and no decisions. Start with five.
- Dashboards nobody opens. Push alerts to owners instead of waiting for them to look — and send an alert only when someone must act.
- No owner. A signal that belongs to everyone is acted on by nobody.
- Automating before the habit exists. If the loop does not work by hand, automation will not save it.
- Agents deciding too early. Recommendations first; decisions later, on evidence.
Evaluation
After a month or two, ask:
- Are the five decisions made on time more often than before?
- How many problems were caught by a signal rather than by a customer?
- Has the weekly meeting stopped arguing about which number is right?
- Which signals fired and were ignored? Remove or fix them.
- Did the actions change the outcomes they were meant to protect — fewer stock-outs, faster payment, fewer lost customers?
Security and governance
Small teams often share logins and export data freely. Give each person their own account, with access only to what they need. Keep personal data out of shared spreadsheets where you can, and collect only what the decisions require. If you add an agent, give it read-only access first, its own identity, and a record of what it did. Decide who may change the register's thresholds.
Cost considerations
Most of the cost is time: a few days to set up the register and the numbers, then thirty minutes a week to run it. Use the tools you have; most accounting, sales and shop systems can export or connect what you need. Spend money only where a specific step is slow or error-prone. If your data is scattered and you do not know where to start, the Guild's Data Landscape Audit maps sources, gaps, owners and decision needs.
Example
Illustrative example, not a client engagement.
A twelve-person kitchen-fitting business runs on an accounting package, a CRM, a shared calendar and a spreadsheet of jobs. Cash is tight because invoices are chased irregularly, and materials are sometimes ordered too late for an installation.
The owner writes a register with four decisions: chase overdue invoices, order materials for jobs starting in three weeks, follow up quotes not answered in ten days, and review job margins monthly. Alerts go to the office manager and the installation lead. At the end of each month the review asks two questions: were invoices chased on the day the register says, and did any job wait for materials? Only when both answers have been "yes, on time" for a while does the next step — an agent that drafts quote follow-ups for approval — make sense.
Checklist
- Five recurring decisions are written down, each with one owner.
- Each decision has a signal in an existing system and a clear condition.
- Each has a default action and a deadline.
- Owners are alerted when conditions are met, and only then.
- A thirty-minute weekly review checks every decision.
- Signals nobody acts on are removed.
- Outcomes are measured, not only actions.
- Everyone has their own account; personal data stays out of shared sheets.
- Any agent starts read-only, with its own identity and a record of its actions.
