Ask most businesses whether their team uses AI and the answer is yes. Someone drafts emails with ChatGPT. Someone else summarises meeting notes. A few people have tried Copilot or Gemini. Ask whether AI has taken any real work off the team, the weekly chasing, reconciling and reporting, and the answer is usually no.
That gap isn't about the quality of the models. It comes down to where the AI sits, what it can reach, and who has decided what it's allowed to do.
AI on the side
In most companies, AI lives in a browser tab. People copy something in, get something back and paste it somewhere else. That's useful for writing and thinking, and it's a fair place to start. But it means three things.
It can't see the work
The information a business runs on sits in its systems: the accounts in Xero or QuickBooks, the pipeline in HubSpot or Pipedrive, payments in Stripe or GoCardless, tickets in Zendesk or Intercom, projects in Asana or Jira, and a lot of email. A chat assistant on the side sees only what someone pastes into it.
Assistants are getting connectors. Microsoft 365 Copilot can work with a person's own email, files and meetings, Gemini does something similar inside Google Workspace, and ChatGPT's business plans can link to some workplace apps. That helps one person find and summarise things. It rarely reaches across the accounts, the CRM and the payment system at once, and it works with each person's own access rather than rules the business has agreed.
Nobody has set the rules
Even where a tool could be connected, most owners hesitate, and for good reason. Who would it act as? Could it send an email to a customer? Could it change an invoice? If nobody can answer those questions, the safe choice is to keep AI away from anything that matters. So it stays on the side.
Everyone is on their own
Each person works out their own prompts and their own habits. What works for one finance assistant doesn't reach the rest of the team, nothing is written down, and when that person leaves the know-how leaves with them. There's no shared view of what AI is being used for or what it costs.
What using AI properly looks like
Using AI properly means agents that can reach the systems a job needs, act within rules you've set, and be called by the team from the apps it already uses. Here is what that looks like in a typical growing business.
Chasing
Every Monday, an agent finds invoices past their terms in Xero, checks the last reminder sent from the shared inbox, and sends the next one in your tone. After the second reminder it hands the account to a named person instead of sending a third. The same pattern covers quotes with no reply after a week, timesheets not yet submitted and contracts waiting for signature.
Reconciling
Each morning, an agent matches Stripe or GoCardless payouts against orders or invoices and lists anything that doesn't match: a refund with no credit note, a fee that looks wrong, a payment with no customer. Someone still decides what to do about each one, but they start from a short list rather than two spreadsheets.
Reporting
Before the Monday meeting, an agent pulls open deals from HubSpot, recent emails and calls, and this month's invoiced revenue, and posts a summary to the leadership channel. Anyone can ask a follow-up question with @ in Slack or Google Chat, such as "which of these deals haven't had a reply in two weeks?", and get an answer from the same systems.
Handing over
When a deal is marked won, an agent creates the project in Asana, sets up the customer in the accounting system, drafts the welcome email and tells the delivery lead what was promised in the sales notes. Nothing gets lost between sales and delivery because nobody has to remember to copy it.
None of these jobs is glamorous. Together they're often a large part of an operations, finance or account manager's week.
How to start
The businesses that get value from AI quickly tend to start small and specific.
Pick one team
Choose the team with the most routine work that crosses between systems, often finance, operations or sales support. One team gives you a clear owner, a clear before-and-after, and people who will tell you honestly whether it helps.
Pick one recurring job
Look for a job that happens every day or every week, follows a recognisable pattern, and needs information from two or more systems. Avoid jobs that depend mostly on judgement or relationships to start with. Pick something where a mistake would be easy to spot and cheap to fix.
Set the rules first
Before anything is built, agree:
- which systems the agent may read, and which it may change
- who can call it
- what it must never do on its own, such as send money, delete records, or email a customer without review
- what waits for a person's approval, for example a refund over a set limit
- who reviews its work, and how often
Writing this down takes an hour. It's the difference between an experiment people trust and one they quietly stop using.
Measure it
Note how long the job takes today, how often it goes wrong, and how long it takes to fix when it does. After three or four weeks, look again, and include the time spent adjusting the agent. If it isn't clearly better, change the job or drop it.
Working out what it's worth
A rough estimate helps you decide where to start. Count the people who spend part of their week on routine admin that crosses your systems, what an hour of their time costs, and what share of that work an agent could realistically take on. Be honest about the last figure. Most jobs can't be handed over entirely, and agents need looking after.
Work it out for your business
Costs you today
£5,200
a month · £62,400 a year
Comes back to you
£1,820
a month · £21,840 a year
Team hours back
61
a month
Counts recovered team time only, after upkeep but before set-up, subscription and AI usage costs. It leaves out faster answers, fewer mistakes and work that simply wasn't getting done. Use a share you've checked against a normal week of your own work, not a headline figure.
Treat the result as a ceiling, not a forecast. It counts time only, before what the service and AI usage cost. The point is to compare jobs with each other, and to see whether the size of the prize justifies the effort of setting it up.
The control question
The more you connect, the more control matters. An agent that can read your accounts and send email can do a lot of good, and a lot of damage if something steers it wrong. Agents act on what they read, so a hidden instruction in an email or a shared document could try to direct one. That risk is new, and it's real.
The answer isn't to keep AI away from your systems. It's to set it up the way you would for a new member of staff:
- Least access. Each agent reaches only the systems and actions its job needs.
- Named callers. Only the people you choose can ask a given agent for anything.
- Approvals. Anything outside your rules waits for a person: a payment above a limit, a refund, a message to a large list.
- A record of everything. Every run is logged with who asked, what the agent did, which systems it touched and what it cost, so you can check the work and spot drift early.
With those in place, connecting AI to real systems stops being a leap of faith and becomes something a manager can run. You can read more about how we approach this in the activity log section of our homepage.
Make it a team capability
The last piece is people. Value from AI compounds when it stops depending on one enthusiast. Give a manager the controls to add and remove access, read the log and adjust agents as processes change. Keep a short runbook. Review the list of agents every month, retire the ones nobody uses and add the next job.
That's the shift from AI on the side to AI at work: a few well-chosen agents, connected to the systems the work lives in, running to rules the business has agreed, and looked after by people who understand them.
If you'd like help with the first one, our implementation page sets out how we connect your systems, build the first agents with your team and train the people who'll run them. Or book a strategy call and we'll look at where AI could take work off your team first.