Persistent AI: When the Assistant Remembers, Keeps Working and Learns
Most AI still forgets everything the moment the conversation ends. Persistent AI does not: it remembers your rules, keeps working in the background, and learns from the corrections you make. What persistent AI is, how it already shows up in Business Central, why it is useful, the new risks it brings, and how to govern an AI that never goes home.
Most of the AI people have used so far has the memory of a goldfish. You open a chat, explain your business, get an answer, close the window, and next time you start again from nothing. Every conversation is a first day at work.
Persistent AI is the opposite. It remembers what it has been told and what it has done. It keeps working when nobody is looking at it. And it gets better from the corrections you make. That changes what AI can do in an ERP, and it changes how you have to manage it.
What "persistent" actually means
There are four things that make an AI persistent, and most real systems have some of them rather than all:
Memory. It keeps context between sessions: your rules, your preferences, the decisions you have already made, the exceptions you have explained once.
Continuity. It runs in the background, watching for work rather than waiting to be asked. A mailbox, a queue, a schedule.
Learning. It adjusts from feedback. When you correct how it coded an invoice, it codes the next one from that supplier the way you showed it.
Identity. It exists as its own user, with its own permissions and its own history, rather than borrowing yours for the length of a chat.
A chatbot has none of these. A good human employee has all four. Persistent AI sits somewhere in between, and moves closer to the employee every year.
How it already shows up in Business Central
You don't have to wait for the future to see this. Business Central's agents are persistent in several of those ways already (see agentic AI in Business Central):
They run continuously. The Sales Order Agent and Payables Agent watch a mailbox and pick up work as it arrives, without anyone starting them each time.
They have their own identity. Each agent runs as its own user, with its own permission sets, so what it can touch is controlled like any other user (see permissions).
They keep a history. The agent task pane shows what each agent is working on and what it has done, and the work it creates sits in Business Central with a normal audit trail.
They learn from corrections. The Payables Agent learns from the mapping corrections users make, so a supplier's invoice that needed fixing once is handled better next time.
External agents can work against live data. Through the Business Central MCP server, an agent built in Copilot Studio, or an assistant such as Claude with its own memory, can work against your current data with BC's permissions, rather than a copy from last week.
And the underlying data in Business Central is, of course, the most persistent memory any business has: every posted document, every ledger entry, every change log entry. A persistent AI that reads it properly doesn't need to remember your stock levels; it can look.
Why it's useful
You stop repeating yourself. Explain the tolerance for price differences once, not every Monday.
Consistency. The same rule applied the same way at 7am and 7pm, on the first invoice of the month and the five-hundredth.
It notices patterns over time. A supplier whose invoices drift above the PO price, a customer whose orders always arrive on a Friday afternoon, a recurring exception that is really a process problem. Spotting those needs memory.
Work happens when people don't. Overnight invoices are matched and waiting for review when the AP team arrives.
It compounds. Every correction makes the next run better, which is how the human-in-the-loop review gradually needs less of the human.
The new risks
An AI that remembers and keeps working brings risks that a forgetful chatbot never did. None of them are reasons not to use it. All of them need managing.
It remembers the wrong thing. If someone corrects an agent badly, it learns the mistake and repeats it confidently. A one-off error becomes a pattern. Corrections need the same care as the original setup.
Memory goes stale. A remembered rule ("this supplier always invoices in euros") can be overtaken by events. Persistent AI should check live data in Business Central rather than trust what it remembers, and its memory needs reviewing.
What it remembers is data. If an agent's memory holds customer names, employee expenses or bank details, that is personal and sensitive data, with all the usual obligations: where it is stored, who can see it, how long it is kept, and whether it is covered when someone makes a subject access request (see data classification and GDPR).
Drift. An agent that worked well in March can slowly get worse as suppliers change layouts, new items arrive and processes change. Nobody notices until the error rate is high. Accuracy has to be measured over time, not once at go-live.
Running costs that never stop. Something that runs all the time is charged all the time. Agent work in Business Central is billed on consumption, so a persistent agent is a monthly running cost to budget and watch (see the real cost of AI agents).
Access creep. Agents, like people, collect permissions over time: "can it just also do this?". Six months later it can do far more than its job needs.
Accountability. When an agent that has been running for a year does something wrong, someone has to own it. "The AI did it" is not an answer an auditor will accept.
Governing an AI that never goes home
Treat a persistent agent the way you would treat a long-serving member of staff who works nights unsupervised: trusted, but managed.
Give it an owner. A named person responsible for what it does, who reviews its work and its corrections.
Keep permissions narrow, and review them. Its own permission set, scoped to its job, checked quarterly like any user's.
Keep a human on the posting. Agents prepare, people approve, until accuracy has been proven over months, not weeks.
Review what it has learned. Periodically check the rules and mappings it is applying, and remove the bad ones.
Measure it every month. Volume handled, error rate, how often people override it, cost per transaction. Watch the trend, not the snapshot.
Know what it stores, where, and for how long. Apply the same retention and data protection rules you apply to everything else.
Keep an off switch and a manual process. If the agent stops or misbehaves, the business must still be able to pay its suppliers by hand. Keep the skills in the team.
Use the audit trail. Business Central and Microsoft Purview log what agents do. Look at the logs, rather than just knowing they exist.
Where it's heading
AI is moving steadily from tools you pick up to colleagues that stay. Expect agents that remember more about your business and its exceptions, that run for longer without supervision as they prove themselves, that work alongside other agents across Business Central, Outlook and Teams, and that are measured and audited like any other part of the finance function. The businesses that get the most from it will be the ones with clean data, written-down processes and good governance, because a persistent AI learns from whatever it is given. Give it a mess and it will remember the mess.
Where to start
Pick one continuous, high-volume process (supplier invoices are usually best), give the agent its own narrow permissions and a named owner, keep a person approving everything, and measure it monthly. Let it earn its persistence. (See what the Payables Agent actually does.)
If you'd like help setting up an agent that stays useful a year from now, not just on day one, book a 30-minute call.
Email this post as a PDF
A branded PDF of this post, sent to your inbox to keep, print or share. Copy in colleagues if you like.
We only use these addresses to send this PDF. Nobody is added to a mailing list. Mike at Business Boxed receives a copy. See our privacy notice.
