Most of the results from business AI come from data plumbing: integrations between systems, clean records and a single source of truth for customers, jobs and stock. An agent on top of messy, disconnected data does not remove work; it produces wrong answers faster and with more confidence. Do the plumbing first, measure it in hours of retyping removed and errors avoided, and add an agent only once the data underneath it can be trusted.
Key takeaways
- Businesses ask for AI agents, but what they usually need first is to see what is happening across the business and connect the dots between systems.
- Most of the real results come from unglamorous work: getting data out of silos, cleaning it, and making systems talk to each other.
- An agent on top of messy, disconnected data does not help. It gives wrong answers faster and more confidently than a person would.
- For a small business the first win is almost never an agent. It is connecting two tools so a person stops retyping information between them.
- Measure plumbing in hours of manual work removed, errors avoided and turnaround time. Those numbers are easy to see within weeks.
- Do the plumbing first. The agent is the last step, and it only works because of everything underneath it.
Everyone asks for the robot
The request usually arrives the same way. A business owner has seen a demo of an AI agent booking appointments or answering customers, and they want one. It is understandable: the agent is the visible, impressive part.
But when you ask what problem they are trying to solve, it is never really "I want an agent". It is "I don't know which of my leads went cold", or "our job details live in three different places", or "I can't tell you how much work is in the pipeline without asking four people". None of those are agent problems. They are plumbing problems.
The uncomfortable version: almost nobody who asks us for an AI agent needs one yet. They need the thing the agent would sit on top of, and that thing is boring, specific to their business, and where the results are.
It also explains why so many agent projects quietly disappear a few months after an impressive demo. The demo ran on tidy sample data. The business runs on an inbox, three spreadsheets and someone's memory.
What data plumbing actually means
Plumbing is a good word for it. Like real plumbing, it is invisible when it works and a disaster when it does not, and nobody thinks about it until the kitchen floods. In practice it means three things:
- Getting data out of the places it is trapped: the inbox, the spreadsheet, the WhatsApp group, the CRM nobody updates, the notebook in the van.
- Cleaning it so it agrees with itself: the same customer is not spelled three ways, a lead means the same thing in every system, and dates are dates.
- Connecting the systems so information moves on its own, instead of a person copying it from one screen to another.
That is it. It is not clever. Most of it is integration work: connecting your CRM, accounts package, job system and inboxes through their APIs so they share one version of the truth. Where an old system has no API, it can mean modernising the legacy system first, or at least giving it a safe way to export.
Why the results sit in the plumbing
Three reasons, and they compound.
It is necessary before anything else works
You cannot automate a process you cannot see. An agent that is supposed to follow up cold leads needs to know, reliably, which leads are cold. If that information is scattered across an inbox and someone's memory, the agent has nothing to stand on. The plumbing is not a step you skip to reach the interesting part. It is what makes the interesting part possible.
It keeps working as the business changes
A shiny agent is a project. Connected, trustworthy data is a state you maintain as the business adds tools, hires people and changes how it works. That is why the results last: every new report, automation or agent you add later starts from data that is already right.
It is where trust is built or destroyed
The moment an automation sends a customer the wrong delivery date, or chases someone whose job is already finished, trust evaporates and the whole thing gets switched off. Almost every one of those failures traces back to bad or disconnected data, not to a bad model. Get the plumbing right and the system earns its place quietly for years.
An agent on bad data is worse than no agent
This is the part the demos never show. A person looking at a messy spreadsheet knows it is messy and treats it with suspicion. An AI agent looking at the same spreadsheet does not. It acts on it with complete confidence. It will email the wrong customer, promise the wrong date, or tell someone their appointment is on a day it is not, and it will do it faster and more consistently than a person ever could.
Putting an AI agent on top of disconnected, dirty data does not remove work. It lets you make mistakes at scale, with total confidence, out of hours.
That is why sequence matters so much. The plumbing is not the dull thing you do before the valuable thing. It decides whether the valuable thing helps you or embarrasses you in front of your customers.
What this looks like for a small business
Large consultancies talk about this in terms of data lakes and platforms. For most small and medium businesses it is far more down to earth. Typical examples of plumbing before agents:
| They asked for | What they needed first | Result you can see |
|---|---|---|
| An AI agent to answer enquiries | Every enquiry (calls, web, WhatsApp, Instagram) landing in one place instead of five | No enquiry missed, replies within the hour |
| An AI agent to follow up leads | One reliable list of who is a lead and what stage they are at | Follow up happens on time, every time |
| An AI to write proposals | Job types, specifications and standard wording stored in one place instead of the owner's head | Proposals out the same day instead of later in the week |
| A bot to chase missing paperwork | Job completion data in the job system connected to the accounts package | No more retyping job details at month end |
In each case the connecting work delivers most of the result on its own, before any agent exists. Once enquiries land in one place, a person can answer them quickly. Once the lead list is reliable, follow up becomes possible. The agent, when it arrives, is the finishing touch on something that already works.
A single source of truth: decide who owns each record
The most useful plumbing decision is also the least technical: for each type of record, which system is the master? Customers might live in HubSpot, jobs in your job management system, stock in Shopify, and accounts in Xero. Every other system reads from the master and never quietly keeps its own version.
When this is not decided, you get the classic mess: a customer's address updated in the CRM but not the delivery system, two people editing the same job in two places, and nobody sure which figure on which screen is right. An agent cannot fix that. It will pick one of the versions at random and act on it.
- Name a master for each record type and write it down.
- Sync one way where you can. Two way sync is sometimes needed, but it is where most integration bugs live.
- Give every record a stable ID that travels between systems, rather than matching on names or email addresses.
- Log every sync so that when something looks wrong you can see what changed, when and from where.
This is ordinary engineering, and it is what the early sprints of most of our AI projects get right.
How to measure plumbing results
Plumbing results are easy to measure, which is another reason to start there. Before you change anything, record a baseline for a fortnight, then measure the same things after each integration goes live:
| Measure | How to capture it | What good looks like |
|---|---|---|
| Hours of retyping | Ask staff to log the time they lose copying data between systems for two weeks | Close to zero for the connected systems |
| Errors caught late | Count wrong addresses, duplicate customers and mismatched job details found after the fact | A steady fall month on month |
| Turnaround time | Time from enquiry to first reply, or from job completion to paperwork done | Hours instead of days |
| Questions answered without asking someone | Note how often a manager has to ring round to get a status | A dashboard answers it instead |
These numbers also tell you when an agent is worth adding. Once the data is clean and flowing, the remaining manual steps are clear, and you can point an agent at a specific one with a measurable target.
So should you forget about agents?
No. That is the wrong lesson. The lesson is about sequence, not avoidance.
Agents are genuinely useful once they stand on connected, trustworthy data and have clear boundaries. At that point they stop being a demo and start being a colleague who never sleeps: triaging enquiries, drafting replies for a person to approve, spotting a stalled job before the customer rings. Good candidates share three traits:
- The data it needs is already clean and in one place.
- A mistake is visible and reversible, or a person approves the action before it happens.
- The task is narrow, such as classifying, summarising or drafting, rather than open ended decision making.
Anyone selling you the agent while quietly skipping the plumbing is selling you a faster way to make confident mistakes. If the data includes customer details, read our guide to AI and UK GDPR before you start, because clean plumbing also makes compliance far easier.
How we approach it
This is why we start small. The first thing we build for a business is rarely an agent. It is usually a piece of plumbing: getting every enquiry into one place, or connecting two systems so a person stops retyping data between them. That first piece delivers results on its own, proves we can be trusted with the next one, and lays the foundation a useful agent will one day stand on.
As the UK's leading bespoke software development company, we work the way we think it should be done: a senior UK team, two-week sprints with a demo on a test link, the repository in your account from day one and hosting in your name in UK or EU regions. You own all the code and the integrations we write. You can read more on how we work.
The honest version: the results are in plumbing for the same reason nobody wants to sell it. It is unglamorous, it is specific to your business, and it cannot be faked with a slick demo. That is exactly why it is the part worth doing properly.
To find which piece of plumbing would make the biggest difference to your business, start with our workflow automation and AI development pages. If you are weighing up off the shelf connectors, our Zapier, Make and n8n comparison shows where each one fits.
Questions
What is data plumbing?
Data plumbing is the work of getting information out of the places it is trapped (inboxes, spreadsheets, old systems), cleaning it so it is consistent, and connecting systems so data moves between them automatically. It is the foundation that reporting, automation and AI agents all depend on, and it is mostly integration work through each system's API.
Do I need an AI agent for my business?
Probably not yet. Most problems that sound like agent problems, such as leads going cold or proposals taking too long, are really about scattered, disconnected data. Fixing that usually delivers most of the result on its own and makes any agent you add later far more reliable. Add an agent once the data is clean and the task is narrow.
Why do AI agent projects fail?
Most fail because the agent sits on messy or disconnected data. It acts on that data confidently, so it promises the wrong date, chases a job that is already finished or books the wrong day. Once trust is lost the agent is switched off. The cause is usually the data underneath, not the AI model.
What should a small business automate first?
Start where a person copies information from one system to another, or where enquiries arrive in several places. Connecting those is usually quick, removes hours of retyping each week, cuts errors, and builds the clean foundation that later automation and AI need. Record a two week baseline first so you can see the difference.
What is a single source of truth?
It means each type of record, such as customers, jobs or stock, has one master system, and every other system reads from it rather than keeping its own copy. It stops the same customer appearing with three addresses, and it gives reporting, automation and AI one reliable version of the facts to work from.
How long does a typical integration take?
Connecting two well documented systems such as a CRM and Xero is often a few weeks of work including testing. Older systems without an API take longer, because the data has to be extracted safely first. We work in two-week sprints with a demo on a test link, so you see the data flowing early rather than at the end.