What changes when your business system starts doing the groundwork

An everyday look at agentic ERP for companies of 20 to 200 people

In short

In most growing companies, a good part of the week goes on work around the business system rather than on the business itself: typing the same figures in again, matching supplier invoices to orders, chasing customers who pay late, hunting for the report that answers a simple question. None of it is hard. It is constant, and it tends to land on the same few people. In the UK, businesses hit by late payment spend on average 86 hours a year just chasing what they're owed [1].

An agentic ERP is a business system with software helpers, often called agents, working inside it. They read the company's own records, notice what needs attention and prepare the routine work: a short list of what matters this morning, a reminder to a late payer, a restock order, a quotation, an answer to "why did our margin fall last week?" with the report open beside it. A person still checks and approves every document before anything is posted, paid or sent. What changes is where people spend their time: less on gathering and re-typing, more on deciding, talking to customers and suppliers, and running the business.

The savings are real, but ordinary: hours given back every week, cash collected sooner, fewer duplicate or fraudulent payments, stock shortages seen before a customer notices, work billed that would otherwise slip through. They build up gradually, a few routines at a time, and they depend on reasonably tidy records and on people who stay in charge. The rest of this article goes through each part of the business: what changes, what stays the same, and how to tell whether it's working. Each part ends with where Dynocortex is different, both what works today and what's planned.

1. Where the week goes today

Every business is different, but the same patterns show up in study after study:

  • Getting paid takes chasing. UK businesses are owed an estimated £26 billion in late payments at any one time, about £17,000 for each business affected, and 28% of businesses are affected each year [1].
  • Paying suppliers takes handling. Processing a single supplier invoice costs the most efficient quarter of organizations $1.42 and the least efficient quarter $6, according to the benchmarking group APQC [2]. Even the best performers lose about 0.8% of what they pay out to duplicate or wrong payments; for the weakest it's 2% [3].
  • Fraud targets the payment run. In a 2026 survey of US finance professionals, 74% of organizations had faced attempts to trick them by email into paying the wrong party [4]. In the UK, criminals posing as suppliers with "new bank details" took £41.3 million in 2025, two thirds of it from business accounts, even as those losses fell to their lowest level yet [5].
  • Closing the month takes days. In APQC's benchmark of about 2,300 organizations, the median monthly close took 6.4 calendar days, and the slowest quarter took ten days or more [6].
  • Work slips through unbilled. Professional services firms lost an average of 4.5% of revenue to leakage in 2025, in SPI Research's benchmark [7].
  • The people who do this work are hard to find. The US expects about 115,300 openings for accountants and auditors every year to 2035, mostly to replace people leaving the job [8].

These figures describe the problem, not what any one product achieves. But they show where the hours and the money are.

Where Dynocortex is different

  • It starts where the hours go. Our plan takes the costliest routines first: collections and the checks before each payment run, then supplier invoices and the month-end.
  • The whole team gets it. Many business systems charge for each person who logs in, and charge again for each piece of AI work. So the help often reaches only the few people with a license. Dynocortex is priced per company, so the bookkeeper, the buyer and the warehouse lead all get it, not just the managers.

2. What "agentic" means, in plain words

Strip away the buzzwords and an agentic ERP does three things a traditional system doesn't:

  1. It answers. You ask in plain words, "who owes us money past due?" or "where is this order?", and get the answer from the company's own reports, with the report or document open beside you so you can check it.
  2. It watches. It notices what you'd want to know without being asked: a best seller about to run out, a supplier invoice far larger than usual, a good customer suddenly paying late.
  3. It prepares. It drafts the routine documents that follow: the restock request, the purchase order, the quotation, the reminder, the checklist.

And there is one thing it doesn't do: decide. Every document it prepares waits for a person, who sees exactly what will be created and approves it, changes it or throws it away. If something changes before approval, such as a colleague updating a price, the system asks again rather than going ahead with an outdated version.

Think of it less as a robot and more as a very well-organized colleague who has read every record in the system, never gets tired of routine, and always brings the paperwork for you to sign rather than signing it themselves.

Where Dynocortex is different

  • Today: the agents never submit, pay, cancel or delete anything. They only prepare drafts. Some products now let their agents create and change records directly, or approve on their own. Ours don't.
  • Today: each approval is tied to exactly what the card showed: the lines, prices, totals and party. If any of that changed before approval, the approval is refused and the card is shown again. Few products say they check this.
  • Planned: an option so that the person who asked for a document can't also approve it, and an agent never approves its own work.

3. A different Monday morning

Today, Monday often starts with running four reports and working out which of a dozen small fires matters most.

With an agentic ERP, the week starts with a short brief: cash this month, what customers owe and how late, what's due to suppliers, last week's sales. Below it, a handful of items that need someone: an overdue invoice, an item about to run out, a budget line over. Ask "which of these should I deal with first, and who owns each one?" and the list comes back ranked by money at risk, each item with the colleague who holds that decision.

The change isn't dramatic, but it adds up. The first hour of the week goes on acting, not assembling.

Where Dynocortex is different

  • Today: the brief comes to you every morning, from the business system's own reports, with totals only. Below it is a "Needs you" list ranked by money at risk, each item with the colleague who decides. Most business systems give you a dashboard to go and look at. A brief that comes to you is still rare.

4. Money coming in

Collections. Most companies chase the oldest invoice first, out of habit. An agentic ERP looks at what is actually unusual: the customer who always pays at 60 days is less of a worry than the prompt payer who is suddenly 30 days late. It ranks the calls, shows the reminders already sent and any dispute open, and drafts what to say. The call is still yours. Over a few months, that tends to mean cash in sooner and fewer awkward surprises.

Work done but not billed. Shipments that never became invoices, billable hours that never reached a bill, a milestone date passed without an invoice raised. These are easy to miss because nobody owes you that money yet, so it never appears in what's overdue. A system that compares what was delivered and worked against what was billed, every week, catches it while the customer still remembers the work.

Cash ahead. "Will we have enough cash over the next quarter?" usually means a spreadsheet rebuilt from scratch. An agentic ERP builds a week-by-week outlook from what's already in the books: invoices due in, bills due out, rent and payroll on their usual dates, each customer's real payment habit. Every line is traceable and every assumption is written down where you can change it. It's an outlook from today's records, not a prediction, and it makes the tight week visible early enough to do something about it.

Where Dynocortex is different

  • Today: collections are ranked by amount, age and each customer's own payment habit, alongside the reminders already sent. The system finds deliveries and hours that were never billed, and builds a 13-week cash outlook in which every line leads back to a document.
  • Planned: reminder letters drafted for a person to approve and send, promises to pay recorded as follow-ups, and billing drafted from timesheets and contract milestones. Elsewhere, cash outlooks are mostly still in early testing, and automatic reminders often go out with no one checking them first. Billing from contracts is rarely offered at all.

5. Money going out

Supplier invoices. An agentic ERP checks each invoice against what was ordered and what actually arrived, and holds anything billed above the delivery or the agreed price for someone to look at. That turns line-by-line checking into reviewing a short list of exceptions.

Duplicates and fraud. Duplicate invoices rarely look identical: the same supplier, the same amount, a few days apart, an invoice number with one digit swapped. A request to "update our bank details" arrives by email just before a payment run. A system that looks for these patterns before every payment run, and asks a person to confirm a changed bank account by phone, protects the business at the moment that matters most.

The work of paying suppliers doesn't disappear. It becomes checking, instead of searching.

Where Dynocortex is different

  • Today: ask beside any supplier invoice whether it matches its order and what was received, or whether it looks like a bill already paid. An alert also flags an invoice far above that supplier's usual size.
  • Planned, among the first routines we build: a hold before each payment run on likely duplicates and on any supplier whose bank details changed since the last payment, for a person to confirm by phone. The check knows only that the details changed. It never reads the account numbers themselves.
  • Honestly: several products already read supplier invoices from email. We don't yet; it's on our plan. Some of those products only warn about a mismatch and don't hold the payment. We'd rather hold it.

6. Closing the month

A late close is rarely one big problem. It's a dozen small ones found one at a time: a bank line nobody matched, a supplier invoice left in draft, goods received but not yet billed, a department over budget with no explanation.

An agentic ERP runs those checks in the last week of the month, not after it, and gives each one an owner and a date. It can draft the first version of the monthly commentary, what moved and why, from the invoices behind each line, for the controller to check. The accountant still closes the books. They just stop discovering the blockers on day five.

Where Dynocortex is different

  • Today: before the month ends, the system lists what will hold up the close, each item with an owner. It checks that the books agree with themselves, and it warns when a stock revaluation stalls, a problem that otherwise leaves stock values wrong without telling anyone. Bank statement lines never reach the AI.
  • Planned: a close checklist with a task for each owner, and draft accruals for the accountant to check. We're building narrow checks a person can trust, not one big "close agent". Some systems are shipping the autonomy first and the controls afterwards.

7. Stock and suppliers

Seeing a shortage coming. Stock on the shelf can look healthy while customer orders have already claimed all of it. An agentic ERP watches projected stock, what's on hand plus what's ordered minus what's promised, against each item's real sales and the supplier's real delivery times, and says so the morning a best seller dips below where it should be. It suggests a quantity, names the supplier, and drafts the order for the buyer to approve. Studies of retailers have long found that items missing from the shelf cost a typical store several percent of its sales [9]. Wholesalers and distributors feel the same problem as lost orders and rushed freight.

Choosing and chasing suppliers. Is the cheaper supplier really cheaper, once freight, duty, failed deliveries and late shipments are counted? The answer is already in the records; the system puts it side by side. When an order is late, it notices and prepares the follow-up, so buyers spend less time asking "where is it?" and more time on the suppliers that matter.

Where Dynocortex is different

  • Today: alerts when a best seller drops below its reorder level, with restock requests and purchase orders drafted for the buyer. A delivery record for each supplier shows late, short and rejected deliveries. Suppliers can be compared on price after freight, duty and failed deliveries are counted.
  • Today: a supplier's own AI agent can ask about that supplier's own orders and invoices. It's read-only, signed with a key registered for that supplier, and switched off until you turn it on. For most products, this kind of exchange between companies' agents is still only announced.
  • Planned: chasing late deliveries, reading suppliers' order confirmations against the order, and restock quantities that follow demand. Elsewhere, agents that read supplier email are mostly still in preview and often billed per run.

8. Sales and customers

Quotations. A customer emails asking for 200 units. The system drafts the quotation from the price list, the customer's terms and the stock available, and a salesperson checks it and sends it. If the price changed in the meantime, the draft is refused and rebuilt, so nobody quotes yesterday's price by accident.

"Where is my order?" One of the most common calls a sales desk takes. The answer means reading the order, its deliveries and whatever stock or purchase will fill the rest. Asked beside the order, the system gives a date for each line, and a weekly list of everyone waiting on a back order, before they call.

Orders on hold. An order held for credit often sits with nobody aware of it: sales sees an order, finance sees a limit. The system explains why it's held, what payment would release it, and who can decide. It doesn't decide on credit itself; a person does.

Where Dynocortex is different

  • Today: quotations are drafted from the price list and the customer's terms, and refused if a price changed before approval. "Where is my order?" gets an answer from the system's own records, as the person asking, with no hand-off to a colleague. Some sales agents hand exactly this question to a person. For an order held for credit, the system shows why and who can release it.
  • Planned: follow-ups on quotations, a short evidence pack for the person who decides on a credit hold, sales orders drafted from a customer's emailed order, and, later, customers' own agents asking about their own orders.

9. Projects, service, production and equipment

The same pattern repeats across the rest of the business:

  • Projects: which jobs made money, and where the hours went over the estimate, while there's still time to act.
  • Service: which customer issues are open, whether they're inside the promised response times, and whether one product is behind most of the warranty claims.
  • Production: which work orders are late and why, and what a product really costs against its plan.
  • Vehicles and equipment: what's due for service or inspection, and which machine is costing more in repairs than it's worth.

In each case the information already exists in the system. The change is that someone, or something, now looks at it every week and brings the short list to the right person.

Where Dynocortex is different

  • Today: you can ask about projects, service, production, quality and equipment beside the record, from live data, and every area has a guided walk-through on the demo company.
  • Planned: short project updates, overdue job cards, projects set up from a contract, and draft production plans and maintenance for a person to approve. In this part of the market, many agents are still previews or demos, and some run only in the vendor's cloud.

10. The people who use it every day

For the bookkeeper, the buyer, the warehouse lead and the sales desk, the most noticeable change is small: ask, and the screen opens. "Where do I see what a customer owes?" opens the report. Help for the screen you're on sits beside it. New staff learn the system by doing their actual work, not by memorizing menus.

It also changes what the job is. Less re-typing what's already in the system, less hunting through menus, fewer evenings spent on the month-end. With accountants and bookkeepers in short supply, the aim isn't a smaller team. It's a team that can handle a growing business without drowning in paperwork, and that spends its skill on judgment rather than data entry.

Where Dynocortex is different

  • Today: "Ask about this screen" sits beside every screen and knows what's open on it: the document, or the list and its filters. Answers open the report they came from, and every figure says which company, currency and period it covers. The Help tab shows the manual's page for that screen, and guided walk-throughs teach by doing the real task.
  • Why it matters: in some products the AI can't be turned on person by person, and answers rarely show which report they came from. With a price per company, nobody is left out to save a license.
  • Planned: help for everyday staff requests such as leave and expenses, and onboarding for new staff.

11. Staying in charge

The most important part of an agentic ERP isn't what it does. It's what it won't do on its own. A trustworthy system makes these things plain:

  • It sees only what you may see. Each person's questions are answered with that person's own permissions, never more.
  • People approve every document, seeing exactly what will be created. Nothing is posted, paid or sent without a person.
  • There's a record of everything: who asked, what was prepared, who approved it, and whether it was changed afterwards. It's a record an auditor can read.
  • It explains; it doesn't advise. It shows the figures, the arithmetic and who decides. Tax, legal and credit decisions stay with people and their advisers.
  • It can be stopped. One switch pauses it, and the business system carries on working as before.

These aren't just reassurances. In Europe, the AI Act asks companies using AI to keep people able to oversee it, and a system built this way supports that from the start.

Where Dynocortex is different

  • Today: every answer is given with the asking person's own permissions, checked through a signed identity on each request. Payroll, bank accounts, staff records and contact details stay out of the AI's reach unless an administrator decides otherwise.
  • Today: the record of every draft is sealed, each entry linked to the one before, and a copy of each seal is kept in a second system. So a quiet edit, even by an administrator, shows up. Most products keep an activity log; few say theirs can show tampering.
  • Today: for an auditor, one record per draft shows who asked, what was shown, who approved and what became of it. One switch stops it all, and an approval counts only once the card was actually on screen.
  • Planned: keeping the person who asked from approving their own request, spending limits per person and per job, and hosting in an EU or US region of your choice. In some products, AI work may be processed outside your region unless an administrator switches that off.

12. What savings to expect, and how to tell

Honest expectations matter, because the hype around AI is loud. The UK government's 2026 survey found that 32% of UK businesses now use AI, double the year before. Most users saw productivity gains, but "generally modest" ones, and 65% of the businesses not using it said they didn't yet see a clear business case [10]. Analysts at Gartner expect more than 40% of ambitious "agentic" projects to be cancelled by the end of 2027, mostly over cost and unclear value [11].

That's not a reason to wait. It's a reason to start small and measure:

  • Start with two or three routines that happen every week and cost real time: collections, supplier invoice checks, the Monday list, restocking.
  • Measure what you can see: hours spent on those routines before and after, days customers take to pay, duplicates caught before payment, work found unbilled, how often a prepared document is approved without changes.
  • Give it a quarter. The value of an agentic ERP is cumulative: a few hours a week, a few days sooner on cash, a few problems caught early. That is the kind of saving that compounds, and the kind a small business actually feels.

Where Dynocortex is different

  • Today: two measures come straight from the system's own records. The first is how often a prepared document is approved without changes, the plainest sign that the drafts are good. The second is what each question cost in AI usage: a few cents on our demo data.
  • Why it matters: many products sell AI in prepaid credits shared with other services. When the credits run out, the AI stops until more are bought, and usage is hard to predict in advance. Dynocortex shows the cost of each kind of work, with no credits to run out.
  • Planned: a monthly spending limit you set.

13. What it doesn't do

An agentic ERP is not an accountant, a lawyer or a credit committee. It doesn't close the books, file taxes, or decide who gets credit. Its answers come from an AI model reading the company's records, so they can be wrong, which is why every figure says which company, currency and period it covers, and opens the report it came from so a person can check. It is only as good as the records it reads: messy data gives messy answers, and the first weeks often go on tidying up.

Where Dynocortex is different

  • We publish our limits. Today Dynocortex doesn't read invoices or orders from email, match bank lines, change or submit documents, run payroll, or give tax advice. It prepares four kinds of documents: restock requests, purchase orders, quotations and tasks. We'd rather you knew that before you try it than after.

14. How we approach it at Dynocortex

Dynocortex is an open-source ERP, covering accounting, stock, buying, selling, manufacturing and projects, with these agents built in. We built it because we spent years implementing business systems and building AI tools for large companies, and kept seeing the same two things: people buried in routine, and AI that sounded sure while being wrong. Our answer is the one this article describes: let the software do the groundwork, from the company's own records, and keep people in charge of every decision.

Today, Dynocortex answers questions in plain words beside every screen, sends a morning brief and alerts, and prepares restock requests, quotations and purchase orders for a person to approve, with every draft on the record and a switch that stops it all. The other routines in this article are planned, and each part above says which. You can try every one we've built on our demo company, a fictional electrical wholesaler with a year of trade, and see exactly what it does and doesn't do.

It's priced per company, not per user, so it doesn't punish you for giving the whole team access, and you can run it yourself or have us run it for you.

Where Dynocortex is different

  • Open source. You can read the code, run it yourself, and keep using it if you stop paying us. Most competing systems are closed, and some charge extra just to read your own data directly.
  • Your data stays in your system. The agents read your records where they are; they don't keep a copy of your books.
  • One price per company, with AI usage shown per kind of work, rather than licenses per person plus credits.
  • Run it your way: on your own servers or hosted by us, with hosting in EU and US regions planned. Some products' agents run only in the vendor's own cloud.

Dynocortex explains your company's own figures and who decides. It does not give tax, legal, credit or investment advice; for those, ask a qualified adviser.

Sources

  1. Department for Business and Trade (UK), Late payment consultation: time to pay up – government response, updated Jul 24, 2026. gov.uk
  2. APQC data in P. D. Wiggins, "How Much Does It Cost to Process an Invoice?", CFO.com, Dec 7, 2022. cfo.com
  3. APQC data in P. D. Wiggins, "Detect and Prevent Duplicate or Erroneous Payments", CFO.com, Mar 3, 2020. cfo.com
  4. Association for Financial Professionals, 2026 Payments Fraud and Control Survey (465 US practitioners), Apr 14, 2026. financialprofessionals.org
  5. UK Finance, Annual Fraud Report 2026, Jun 2026 (invoice and mandate scams). ukfinance.org.uk
  6. APQC General Accounting benchmarks in P. D. Wiggins, "Cycle Time for Monthly Close", CFO.com, Mar 5, 2018. cfo.com
  7. SPI Research, 2026 Professional Services Maturity Benchmark, as summarized by Deltek, Jul 15, 2026. deltek.com
  8. U.S. Bureau of Labor Statistics, Occupational Outlook Handbook: Accountants and Auditors, 2025–35 projections. bls.gov
  9. T. Gruen, D. Corsten, S. Bharadwaj, Retail Out-of-Stocks: A Worldwide Examination of Causes, Rates, and Consumer Responses, Grocery Manufacturers of America, 2002.
  10. Department for Science, Innovation and Technology (UK), AI adoption survey report 2026: executive summary (3,503 interviews), Oct 29, 2026. gov.uk
  11. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", Jun 25, 2025. gartner.com