White paper · All distributors
The questions that cost distributors most: how AI agents give UK distributors their time back
Stop answering the same questions twice. A practical guide to the repetitive questions inside a distribution business, what an AI agent can safely take on, and how to start with one workflow.
Executive summary
Your business already has the answers. Finding them is the job nobody budgets for.
Every distributor answers the same questions every day. Who has gone quiet? What needs reordering? Which invoices are overdue? Will this product suit my application? The answers exist, but they sit across ERP exports, spreadsheets, catalogues, inboxes and the heads of two or three experienced people.
In discovery calls with UK distributors and suppliers in 2026, the pattern was consistent. Owners told us their real problem is not a lack of data. It is that getting a straight answer out of it takes a person, a spreadsheet and an afternoon. One director described the company’s reorder process as being done “manually, which is unbelievable”. Another put it simply: the priority was “understanding data and sort of access to what we need”.
AI agents are now practical for businesses of 10 to 25 people. A quarter of UK businesses reported using some form of AI by late 2025, up from 9% two years earlier1, and the cost of running a capable model has fallen more than 280-fold since late 20222. What has not changed is the need for answers a team can check.
This paper explains, in plain English, what an AI agent is and what it is not. It describes two ways to use one: an internal agent that answers your team’s questions about sales, stock, credit, customer health and products, and a website agent that answers customers’ product questions and hands over to a person when it should.
It then sets out a simple business case method, answers the objections we hear most often, and offers a readiness checklist. The approach throughout is the same: start with one workflow, on the information you already have, read-only, with the source shown on every answer and the checks agreed before anything goes live.
01 · The cost of the same question twice
The questions that arrive every day
Most owner-run distributors of 10 to 25 people run an ERP such as SAP, Sage or Business Central, plus a set of spreadsheets that make the ERP usable. The business works. But a surprising share of each day goes on answering questions whose answers already exist somewhere.
None of these questions is hard. Each one is a few minutes of opening an export, filtering a sheet, checking a note and asking a colleague. The cost is in the repetition, and in who ends up doing it: usually the people who know the business best, and whose time is most valuable on the phone to customers.
The managing director of a UK tools supplier described the arithmetic plainly. His reps were spending around three hours a day talking to customers. The rest went on preparation, searching and admin. “If I get 3 hours talk time out of a person who’s doing 8 hours work,” he asked, what are they doing with the other five?
- 19%of working time spent by interaction workers trying to track down the information they need to complete tasks. McKinsey Global Institute, 2012, using IDC data3.
“Internal, I would have said: understanding data and sort of access to what we need.”
| Area | The question, as it is asked | Where the answer sits today |
|---|---|---|
| Sales | Who hasn’t reordered? What did this customer buy last time? Which accounts are spending less? | Sales history export, CRM notes, a rep’s memory |
| Stock | What is running low? What should we reorder, given lead times? What hasn’t moved? | Stock and purchase exports, picking lists, the buyer’s own rules |
| Credit | Which invoices are overdue? Who is paying later than usual? What did they promise last time? | Aged debt report, a credit controller’s spreadsheet of conversations |
| Products | What is the closest alternative? What specification does this code have? | Catalogues, data sheets, supplier PDFs, a technical colleague |
| Customers | Will this work for my application? Do you have it in this width? Can I get a quote? | Website pages, phone and email, the contact form |
Knowledge in people’s heads, data trapped in exports
Two things make the repetitive questions expensive. The rules live in people, and the data lives in exports.
The rules live in people. Every distributor has rules that are not written down. These two film codes are the same product from different suppliers. That customer always pays late in August and it is fine. Ignore stock that arrived last week when you look for slow movers. Sales figures are always ex-VAT. These rules make the difference between a right answer and a misleading one, and they usually sit with one or two long-serving people. When that person is on holiday, busy, or retiring, the answer slows down or stops. We spoke to several owners in 2026 who were planning a handover to the next generation or a new owner. The knowledge question is real.
The data lives in exports. Most ERP systems in this size of business are good at recording transactions and poor at answering questions. So the team exports. A custom query becomes a monthly spreadsheet; a VLOOKUP breaks when a customer changes its name; a reorder decision is made line by line. A credit controller keeps years of payment conversations in a sheet of her own.
- ERP exports
- Spreadsheets
- Catalogues and data sheets
- Inboxes
- CRM notes
- People’s heads
- One person, several files and an afternoon
“It is bonkers that we are still literally in the Stone Age of having to say, right, we’ve got a hundred of these... we now need to order another 50.”
“If we started off at the moment, we would end up with getting a lot of garbage out. Because it’s garbage in at the moment.”
02 · Why now
Three things changed at once
For years, software that could read a spreadsheet, follow a company’s rules and answer in plain English was out of reach for a 15-person business. It is not any more. Three shifts explain why.
Source: ONS1. Official statistics in development. Among businesses with 250 or more employees the figure was 44% in late December 2025.
- 280×Fall in the cost of querying an AI model that performs at GPT-3.5 level, between November 2022 and October 2024. Stanford HAI, AI Index Report 20252.
- 75%of B2B buyers say they prefer a rep-free sales experience. Yet buyers who complete a self-service digital purchase are 1.65 times more likely to regret it. Gartner B2B Buying Report, 20234.
1. The cost collapsed
Running a capable model is now cheap enough that the economics depend on the build and the upkeep, not on the AI itself. That puts private agents within reach of owner-run businesses.
2. Your peers have started
On our own calls in October 2026, one distributor already had AI connected to its ERP for quotes and purchase orders. Another was finishing a website agent “to reduce the time actual people spend replying to questions”.
3. Buyers expect answers
Trade customers want to self-serve, but they also want to get it right. The opportunity is fast, sourced answers with a clear route to a person, not a contact form or a guessing chatbot.
03 · What an AI agent is, in plain English
A colleague who has read every export, and shows their working
An AI agent is software that reads the information you give it, follows the rules you write down, and does a defined job: answering a question, preparing a list, drafting a reply. Then it shows you where the answer came from. Unlike a general chat tool, it does a narrower job, on your information, in a way your team can check.
1 · Your information
ERP exports (SAP, Sage, Business Central), Excel and CSV, catalogues and data sheets, CRM records and team rules.
2 · The agent
Reads the agreed sources, applies your written rules, answers in plain English and logs every question.
3 · A checkable answer
The answer itself, the file and rows it used, the rule it applied, and a flag when it is unsure.
4 · Your person decides
Checks against the source, corrects it once (saved as a rule) and takes the action.
What a well-built agent does
- Shows the source on every answer: the file, the rows, the rule.
- Applies team rules, written down once and used for everyone, such as “sales are always ex-VAT” or “ignore stock received in the last 14 days”.
- Is read-only by default. It needs no ERP write access. Any action is switched on by you and logged.
- Says when it does not know, and hands over to a person.
What it is not
- Not a new ERP. It sits on top of the systems you already run.
- Not another login to manage. Socraty is an implementation partner: the agent is scoped, built, tested and handed over with your team.
- Not a replacement for judgement. It prepares the answer; your people make the call.
- Not trained on your data for anyone else. Your information is kept separate and never used to train AI models.
| What matters on a Monday morning | General AI chat, as it comes | Private agent on your exports |
|---|---|---|
| Knows your product codes, families and units | Only what you paste in each time | Mapped and signed off by your team |
| Shows the exact file, row or rule behind an answer | Not by default | On every answer |
| Remembers a correction | Per person, per chat | Saved once as a team rule, for everyone |
A general description of default behaviour in 2026. Plans, settings and features of general tools vary and change.
04 · The internal agent
One place for the team to ask about sales, stock and credit
The internal agent is where most distributors should start. It answers the questions your own team asks every day, from a rep’s quick account check to the director’s morning view of the business.
It works from what you already have: sales and stock exports from the ERP, the spreadsheets the team maintains, product documents and the rules your people carry in their heads, written down once.
Internal agentExample data
Which accounts are late against their usual ordering interval this week, and what should I check before calling?
Two accounts are later than usual. Customer A is 84 days on against a usual 32. Customer C is 21 days on against a usual 7. Customer B is within its interval.
Team rule 14 says to check for a partial return in the last 30 days first. Customer A has a credit note on 12 August for two damaged units.
Source Sales history by customer, to 31 August · rows 118 to 142
Source Sales analysis (Aug) · credit notes, row 7
Rule Team rule 14: partial returns, saved by Sales
What owners asked for on our calls
- A ranked call list by 8am. The MD of a UK tools supplier wanted reps to arrive to a list of who to call and why, built from the ERP and CRM, with a red or green health view for each account.
- Reorder suggestions that follow the buyer’s own logic. One packaging distributor’s rule was six months of demand, halved, less free stock and stock on order, with equivalent supplier codes read together.
- A credit log that remembers. A credit controller wanted to record each conversation in the agent and ask for a customer’s payment history and usual excuses.
- Lapsed customers spotted early. Asked how they identified customers who had stopped ordering, one distributor said: “We have no process in place.”
“Just an agent that could look at 4 different spreadsheets and say, hey, use XYZ today.”
| Area | Typical questions |
|---|---|
| Sales | Who hasn’t reordered against their usual interval? What could this customer buy but doesn’t? |
| Stock | Which lines will run out before the next delivery? What should we reorder, given lead times? |
| Credit | Which accounts are paying later than usual? What did they say last time we chased? |
| Customer health | Brief me before I ring this customer: what do they buy, what is open, what has changed? |
| Products | What is the difference between these two codes? What is the closest alternative? |
05 · The website agent
Product questions answered while the customer is still interested
Trade customers describe what they need in their own words. The website agent finds the right product from your approved catalogue and hands a qualified enquiry to sales.
On our calls, the usual starting point was a contact form or “we encourage customers to call us directly”. That works when someone is free. Out of hours, or when the technical person is on another call, the question waits. An agent can answer the detailed questions that arrive overnight, so the team starts the morning with the requirements already captured and can send a proper quote.
Ask our product teamExample data
I need shrink film for a semi-automatic machine, about 500mm wide. We use 19 micron now.
Two options in the catalogue fit a 500mm semi-automatic machine at 19 micron. What is your monthly volume?
Around 1,000 rolls.
Thank you. I’ve passed this to the sales team with your requirements so they can confirm price and availability.
Source Catalogue entries for 500mm shrink film, 19 micron
- Enquiry sent to sales
- Width500mm
- Gauge19 micron
- VolumeAbout 1,000 rolls a month
How a good website agent behaves
- Answers only from content you approve: catalogue pages, data sheets, agreed FAQs. It names the entry it used.
- Declines to guess. If the catalogue is silent on outdoor use, it says so and offers a person.
- Hands over with context, so sales gets the width, gauge and volume, not a bare “contact us”.
- Keeps prices and stock with your team unless you decide otherwise and scope it.
- Logs every question, so repeated unanswered questions show where the catalogue needs work.
“We have a pretty responsive team... it would need a person involved, and we’ve got a good team.”
06 · Customer story
A UK packaging distributor: one place to ask, instead of five places to look
A UK supplier of packaging films and machinery, owner-run, on SAP. Its commercial knowledge was spread across SAP Excel exports, spreadsheets and the people who knew where to look.
The problem
The monthly sales analysis was built from a custom SAP query exported to Excel, then corrected by hand whenever a customer’s name changed and a VLOOKUP stopped matching. Reordering was done product by product. Credit control lived in a spreadsheet of conversations going back years. Which accounts had gone quiet, or how one film was moving across several supplier codes, meant opening several files and asking a colleague.
The discovery
We worked in the office with the team on the tasks they actually did. A boardroom conversation about priorities, bottlenecks and roles was recorded and given to the agent as business context. The field mappings were written down: which column is the invoice number, which identifier never changes, and that sales are always read ex-VAT. So were the exceptions, such as equivalent film codes that must be read together, and recent deliveries that should not count as slow-moving stock.
“When it comes to reordering... we do it manually, which is unbelievable.”
“It will be so good if I could actually just enter all my conversations on this.”
| Sector | Packaging films and machinery |
|---|---|
| Systems | SAP, Excel exports, spreadsheets |
| Built | Internal agent and website agent |
| ERP access | None. Worked from exports the team ran |
What was built, and how it was checked
An internal agent
Chat interface with history, and a selector for which datasets an answer should use. Knowledge bases for products, competitors and business data from SAP exports.
Spreadsheet reading
Multi-tab Excel workbooks and SAP exports, read and reconciled before their figures were trusted.
Source and rules
Every answer names the document, rows or rule it relied on. Stock-code grouping only with the team’s sign-off.
A website agent
Answers product and application questions from the website’s own content, captures the enquiry, alerts the team and hands over to a person. No prices shown.
| Task | Before | With the agent |
|---|---|---|
| Monthly sales analysis | Custom query, export, manual VLOOKUP fixes | Ask in plain English; customer number used as the stable key |
| Reorder decisions | Line by line, by hand | Suggested quantities using the buyer’s written rule and lead times |
| Slow-moving stock | Spotted by eye | Flagged, with recent deliveries excluded |
| Customer buying history | Filter several sheets | One question, with the invoices listed and checked against the source |
| Credit control | Personal spreadsheet of conversations | Planned as a separate agent once sales and stock were stable |
How outputs were checked. Staff compared example outputs with the source reports, including row counts, totals and the exclusions above. When an answer was wrong, the correction was saved as a team rule, so it applied next time and for everyone. Live testing surfaced real issues early: a delivery that had just arrived made a line look like it had not moved, so the rule was changed.
What changed. The change was in method rather than in a headline number, and we do not claim a savings figure for this customer. The agent became a place to ask and check, rather than a replacement for judgement. The team owns the agent and can switch it on or off without us.
07 · Building the business case
Estimate the hours before you estimate anything else
A sound business case for an agent starts with time, not technology. Count the repetitive questions, estimate how long they take, and be conservative about how much of that an agent can take on.
- List the questions. Ask each person for the three questions they answer by hand every week.
- Time them honestly. Include the searching and the checking, not just the typing.
- Apply a recoverable share. Assume the agent removes part of the work, not all of it. A person still checks.
- Decide what the time is for. The value is highest when freed hours go to calls, quotes and customers.
Illustrative model Worked example. All inputs are invented for illustration, for a business of about 12 people. Replace them with your own.
| Question type | People | Minutes a day | Hours a week |
|---|---|---|---|
| Stock and reorder checks | 3 | 30 | 7.5 |
| Account checks before calls | 4 | 30 | 10.0 |
| Credit and overdue prep | 1 | 45 | 3.8 |
| Customer product questions | 2 | 40 | 6.7 |
| Total | 27.9 |
Share the agent handles (assumed): 50%. Hours recovered a week: 14.0. Working weeks a year (assumed): 46. Hours recovered a year: about 640, or about £16,000 at an assumed £25 an hour of staff cost. The figure excludes revenue effects, such as quiet accounts won back or reorders not missed, which are often larger but harder to evidence.
- ≈ £100kOne distributor’s business case estimated around £100k a year in potential savings. It was an estimate built before any work started, not a measured result.
08 · Objections, answered
The questions owners ask before they start
These are the objections we hear most often on calls with UK distributors. They are reasonable questions, and each has a straight answer.
“We already have a system.”
Good. The agent sits on top of it and does not replace it. Built-in AI is often unused or too general for your rules: one fastener distributor told us its CRM’s AI was something they had “not really utilised yet”. An agent built on your exports answers your questions, in your terms, with the source shown.
“We’re not ready. Our data isn’t clean.”
Your data does not need to be perfect first. Start with the export your team already trusts. Records the agent is confident about go straight through; anything it is unsure of is flagged for a quick yes or no. Discovery often shows which data fixes matter, and which do not.
“What about data security?”
Your data is kept separate from every other customer’s and used only to run your agents, never to train AI models. It is hosted in the UK or EU, encrypted in transit and at rest. System credentials are held in a secure vault. Every question and answer is logged.
“Do you need access to our ERP?”
No. You can start with the exports and spreadsheets your team already runs. The agent is read-only by default and needs no ERP write access. A live connection can be scoped later, if it earns its place.
“It would need a person involved.”
Agreed. The agent prepares the answer and your person makes the call. It hands over when the catalogue is silent, and your team’s corrections become rules. The aim is to give good people their time back, not to replace them.
“Can’t we just use ChatGPT?”
General tools are excellent, but they only know what you paste in each time. They do not keep your product codes, account intervals or rules for the whole team, and they do not show the exact row behind an answer by default.
09 · The adoption journey
Start with one workflow. Add the next when you’re ready.
The distributors who get value from agents do not start with a platform. They start with one question their team answers by hand every week.
| Phase | What happens |
|---|---|
| 01 · One workflow, one agent | Pick the question that costs your team the most time, usually in sales, stock or credit. Build the internal agent around it, on the exports you already have. Prove it on your own cases. This is where trust is earned. |
| 02 · A second agent | Add the next workflow or the website agent, reusing the sources and rules already written down. Each new agent is quicker because the groundwork is done. |
| 03 · Agents that share what they learn | Agents read each other’s outputs. The website question log informs stock and catalogue decisions. Sales call lists, supplier emails matched to purchase orders and carrier choice per shipment become possible, each with limits you approve. |
How one workflow goes live
- Discovery. We sit with the team, watch the day-to-day tasks and write down the rules, the exceptions and what a good answer looks like.
- Set-up. The agent is configured on the agreed sources: exports, spreadsheets, catalogue, documents.
- Testing. Tested on your own cases with the people who will use it, including the questions it should decline.
- Acceptance. Checks agreed in writing before the build are run with your team. It goes live when it passes.
- Handover. Written notes, a working handover and an agreed plan for upkeep. You own the agent.
Live in weeks
Most pilots are live within a few weeks, depending on the sources and the checks you set.
Fixed scope, in writing
What the agent does, what it must not do and how it will be checked are agreed before the build.
An implementation partner
Not another login. The agent is scoped, built, checked and handed over with your team.
10 · Readiness checklist
Are you ready for a first agent?
Tick the statements that are true for your business today.
0 of 10 ticked
- 7 to 10Ready for a pilot on one workflow. Bring your question to a 15-minute intro.
- 4 to 6Start with discovery. Agree the first question and the rules behind it.
- 0 to 3Probably not yet. Revisit when the questions start costing more time.
We will also say so when an agent is not the right answer. If a piece of work needs a permanent hire or a larger team, you will hear that in the first conversation.
About Socraty
An implementation partner for UK distributors
Socraty scopes, builds, tests and hands over AI agents with the people who will use them. Discovery is done with your team: sitting with them, watching the day-to-day tasks and writing down the business rules and exceptions that matter. The work this paper draws on comes from that process, and from many conversations in 2026 with owners and directors of UK distributors and suppliers.
- Scope stays honest. One agreed workflow at a time, on information you already have, with the checks written down first.
- Handover is real. Written notes, a working handover and an agreed plan for upkeep. You own the agent.
- Limits are stated. If something is not a good fit, you will be told in the first conversation.
Socraty is the trading name of Socraty Technologies Ltd, registered in England and Wales, company number 16723706. Registered office: 128 City Road, London, EC1V 2NX.
Sources and methodology
Sources
- 1
Office for National Statistics, Business insights and impact on the UK economy. Bulletins of 2 October 2025 (late September 2023: 9%; late September 2025: 23%), 3 July 2025 (late June 2025: 21%), 3 April 2025 (late March 2025: 18%) and 8 January 2026 (late December 2025: 25%; 44% for businesses with 250 or more employees). Official statistics in development. Source
- 2
Stanford Institute for Human-Centered AI, Artificial Intelligence Index Report 2025. The cost of querying a model scoring at GPT-3.5 level on MMLU fell by more than 280-fold between November 2022 and October 2024. Source
- 3
McKinsey Global Institute, The social economy: Unlocking value and productivity through social technologies, July 2012. Interaction workers spend 19 percent of working hours trying to track down information needed to complete tasks. The report bases this estimate on IDC data; it is a global estimate for knowledge workers, not specific to distribution. Source
- 4
Gartner, B2B Buying Report, 2023. 75% of B2B buyers say they prefer a rep-free sales experience; buyers who made a digital self-service purchase are 1.65 times more likely to regret it than rep-led buyers. Based on Gartner’s 2021 B2B Buyer Survey. Source
Methodology notes
- Quotes. Quotes from distributors come from recorded discovery calls and working sessions held by Socraty with UK distributors and suppliers between July and October 2026. They are anonymised and described by role and sector only. Some are lightly edited for readability, without changing their meaning. No quote in this paper has been invented.
- Customer story. The customer is described as a UK packaging distributor and is not named. No savings figure is claimed for that project.
- Example data. Screens, customers, product codes and figures shown in the agent examples are invented for illustration. The worked business case uses illustrative inputs and is not a forecast.
- Business case reference. The figure of around £100k a year comes from one distributor’s own business case, estimated before work began. It is not a measured outcome.