AI Is Ready. Your Legacy Systems and Your Organisation Are Not — What WOW London 2026 Taught Me
I spent two days at World of Workato in London: Partner Day on 19 October and the main conference on 20 October. I expected the story to be about smarter agents. It wasn't. The agents on stage were good enough. In almost every session, what held companies back was something older and harder to change.
Two things, to be exact. The systems of record that agents have to work through, often 10 or 20 years old and customised beyond recognition. And the organisation around them: who decides what an agent may do, who owns the rules, and who notices when something goes wrong.
I have spent 15 years implementing ERP for manufacturers and distributors, and this is what I want CIOs and CFOs of supply-chain-heavy companies to hear. The next 12 months of AI value will not be decided by which model you buy. It will be decided by how ready your core systems, above all the ERP and the CRM, and your organisation are to let an agent act.
During the conference I posted two short notes on LinkedIn: one on Day 1 and one on Day 2. This is the longer version, with the evidence behind them.
Same question, two ways of reaching the system
The slide that stayed with me on Day 1 was a simple test from Workato's technical workshop. An agent got the same prompt twice: "give me a complete account intelligence briefing" on one customer. Same model, same account. Only the connection to the business systems changed.
Scaled to 2,500 users running 10 use cases 20 times a month, the demo put the cost at about $121,000 a month against $17,300. That is a vendor's demo, so treat the exact figures with care. The direction matches what customers said on stage the next day.
On the closing panel, a speaker described a finance team using an ERP vendor's MCP server with about 220 tools. Just loading those tool descriptions took around 30,000 tokens of the model's working memory. During a bank reconciliation, the agent started inventing numbers.
The model was not the problem in either case. The problem was how the system of record was exposed to it. An ERP like Infor M3 has thousands of transactions, and a mature CRM has hundreds of objects and custom fields. Handing all of them to an agent is like giving a new employee the full system menu and no job description.
The Workato team also shared preliminary figures from their Automation Index 2026. 91% of organisations already connect AI agents to systems of record. 70% of those agents connect to two or more of them, and almost two-thirds do not just read data but execute tasks. Agents are already inside your ERP and CRM, or will be soon. The question is through which door.
Legacy is where the journey stalls
The customer stories on Day 2 came from very different businesses. Under the surface, they were all about the same thing: getting new automation through old systems.
TCS opened the afternoon with the most honest description of legacy I heard all week. At one client, most integrations were undocumented. The business rules were hiding in the code, and the people who knew the dependencies had left the company. Before any modernisation, the client first had to find out what it actually had. TCS now uses 75 to 80 specialised agents just to read old integration code and write down the rules it contains. Their warning: copy a decade-old integration as it is, and you carry every old wrong decision into the new platform.
Cisco has about 140,000 named mid-market customers and fewer than 1,000 sellers to cover them. Sellers were juggling 8 to 12 disconnected tools a day. The back end, in the speaker's words, is "anything but simple": Oracle ERP, Salesforce, Snowflake, their own commerce tool and four or five partner portals. Connecting them point-to-point took months per use case. After moving to event-driven orchestration, more than 96% of deal requests are resolved in under two hours, sellers save over 16,000 hours a year, and the team behind the process went from 60+ people to under 25. The others moved to revenue work.
The line I wrote down was this one: "It is absolutely pointless to automate the deal registration if it takes five days to quote that customer." Swap "deal registration" for "order intake" and "quote" for "confirm a delivery date", and you have the situation of most manufacturers and distributors I know.
David Lloyd Clubs is the story that changed how I use the word "legacy". In 2023 they replaced their estate with tier-one ERP, finance, HR and payroll systems. The plan was to simplify. In practice, they had to add layers, because the new systems did not fit how the clubs really work. Their employee app went from very high use to almost none, because it could not get data from Workday, the Fourth scheduling system and the learning platform; two of the three had no ready-made connector. The 12,000 team members became the integration, checking one system against another just to book a holiday. As their IT director put it, if systems cannot trade with each other, someone in between has to carry the data, and that someone is your users, members or hotel guests.
Once they had an integration layer, a proof of value took about five days instead of an estimated three weeks, and they costed tokens per user journey during that proof. When the group bought a hotel business with 135 rooms across four clubs, they integrated it in about three months, without a sandbox. Next on their list is a fix for an audit finding: every year, five or six leavers keep their system access for weeks after leaving.
Wood Mackenzie moved about 55 integrations off MuleSoft and IBM App Connect, around Workday, Oracle Fusion, Salesforce and ServiceNow. Their reasons were cost, key-person risk, vendor dependence and speed. One driver sat underneath all the others on their slide: data quality and governance, described as "laying the foundation for continued AI adoption". Their warning for anyone migrating: unoptimised flows copied over from the old platform burned through a year's task budget within months.
A Workato architecture slide put the pattern most clearly. It sorted the enterprise stack by how fast each layer changes. Models and AI tools change every few weeks. Processes, rules, master data and the systems of record change only as fast as the business itself. The fastest-moving part of your stack now waits on the slowest one. Another slide made the same point from the plumbing side: events, databases, on-premise and legacy systems were the one category a standard MCP gateway cannot reach without a custom-built server.
So legacy is not about the age of the software. A three-year-old ERP can be legacy in this sense. Legacy is any system whose rules are undocumented, whose data is hard to reach, and whose gaps are filled by people. For a supply-chain-heavy company, that is almost always the ERP, the CRM and the processes that run between them.
A permitted action is not a completed outcome
One Day 1 example is worth retelling in full, because it is where legacy systems and governance meet.
A customer disputes a $1,200 charge. An agent handles it in four steps: look up the invoice in the ERP, create a credit memo in the ERP, update the service ticket, and tell the customer it is resolved. A gateway checks every call. All four are permitted and logged.
The outcome is still wrong:
The credit is issued without approval. Company policy says credits over $500 need finance sign-off. No single tool call knew that.
The ticket is still open. The ticket update timed out, and nothing reversed the credit or retried the step.
The customer is told the dispute is resolved. The agent's picture of the process no longer matched reality.
The slide's conclusion was short: governance is not the same as execution. I would add the ERP consultant's version. The rules that make a transaction correct, such as approval limits, reversals and the state of a process, live in your ERP and CRM configuration and in a few people's heads. If they are not written down where an agent's actions pass through, no access policy will enforce them.
The Day 2 keynote demo made the same point from the other side. An onboarding process missed its target because cases kept coming back for rework. On analysis, 118 of 139 returns failed on one of three simple, rule-based checks that needed no judgment at all. The fix was not a smarter agent. It was a better-specified process.
Governance is an organisational job, not an IT checkbox
A quote on the keynote screen, attributed to CIOs, summed up why governance now matters more: "With a person, I worry about one bad action. With an agent, I worry about the same bad action happening 10,000 times before I can stop it."
The closing panel on Day 2 brought together Novuna Business Finance, Skyscanner, dmg media, Deloitte and Workato. Their stories were honest. An early HR chatbot that happily told employees each other's salaries. An assistant that, asked to read an encrypted file, started trying passwords. A Workato specialist who mapped a finance team's bank reconciliation for four hours, then said: "Please stop. However you are doing this, we need to re-imagine this whole process."
I turned the panel into six lessons for my Day 2 post. Here they are again, with what each means for a company that runs on ERP and CRM:
Understand the process before choosing a tool. Design for the most likely 80% of cases. In ERP terms: agree the standard order, return or invoice flow before automating the exceptions.
Put the AI centre of excellence where the business work happens. At dmg media it sits inside editorial, not IT. In a manufacturer or distributor, that means the owners of order-to-cash, procure-to-pay and customer service, not only the ERP and CRM teams.
Govern personal AI and enterprise AI differently. A copilot drafting an email is not an agent posting a credit note. Novuna treats them as two different problems.
Plan the route to company data before people ask for it. dmg media's enterprise architect described what happens with desktop assistants: people are happy with their spreadsheets and PDFs for a while, then "invariably they come back and say, we need that data", meaning the data in SAP. Saying no only grew the security backlog. A controlled rollout of connections gave them a safe way to say yes. Decide the door into your ERP and CRM before people find their own.
Bring the whole company in early. Skyscanner paused work for three days of "AI Unlocked" across the business, and saw a step change in adoption. They also admitted it came about three years into their journey.
Budget for running agents, not just building them. Someone must own each agent, check it, and retire it. Silent failures need systematic tests; forgotten agents keep running and keep costing.
Most companies hand this to a security committee in IT. In a manufacturer or distributor, the questions are business questions, and most of them touch the ERP or the CRM:
Decision | Who should own it |
Which end-to-end processes get agents first | COO and process owners |
Which ERP and CRM data and actions agents may use, and with whose access | ERP and CRM owners with IT security |
Which business rules an agent must never bypass (approval limits, credit holds) | Finance controller |
Consumption budget per business transaction | CFO |
Retiring agents that no one owns any more | Process owner |
One counterweight, from Cisco: their advice was to pace behaviour change before governance. Do not mandate; make the benefit obvious, celebrate the people who use it, then formalise. Governance that only says no does not remove the bottleneck. It moves it to shadow AI.
My takeaway from the panel, as I put it on LinkedIn: readiness isn't a gate you pass before AI arrives. Your people already use it. Getting ready is how the organisation catches up.
Four moves for supply-chain-heavy companies
None of this calls for replacing your ERP or CRM. It calls for treating your core systems, and the people who own their processes, as the centre of your AI programme instead of something to work around.
Fix the process before you automate it. Write down the standard flow, the rules and the exceptions. Treat that specification and its test cases as company assets. Building is getting cheap; a wrong specification is not.
Give agents a small, business-level door into your core systems. Not the full transaction catalogue, but 10 to 20 actions in business language, such as "check credit and availability for this order" or "show open quotes and complaints for this customer". Start read-only, use narrow service accounts, log everything, and add write actions one at a time.
Keep rules and process state outside the agent. Approval limits, reversals and long-running steps belong in the orchestration layer or the core system itself, where they are enforced every time. The agent decides what to do next; it should not be the only thing that remembers what was already done.
Govern from the business, and budget for running. Name owners per process and per agent, set a consumption budget per transaction, and plan migrations as a cost peak of their own. Treat legacy clean-up, such as master data and undocumented customisations, as part of the AI budget rather than a separate IT project.
Four questions before your next AI pilot
If you are planning an agent project for 2027, answer these before you choose a tool:
Which end-to-end process are we improving, and how will we measure it? Days from order to cash, or hours per return, not "tasks automated".
Which ERP and CRM data and actions does the agent need? Read, write or approve, and on whose access?
Which business rules must it never bypass, and where are they enforced? If the honest answer is "in one person's head", start there.
Who owns the agent once it is live? Who checks it, pays for it, and switches it off?
Companies that can answer these will get real value from agents, quickly. Companies that cannot will add to what the panel called the graveyard of POCs.
This seam between core systems and the orchestration layer is where we work at Lightning ERP, with Infor M3 as our home ground on one side and agentic automation on the other. It is also why we published a Workato community connector for M3. If you are working through any of these questions, I am happy to compare notes.
Greg Zajączkowski, Managing Director, Lightning ERP
Comments