AI Can’t Outrun Broken Processes
How organisations are turning agentic AI into measurable business outcomes
Agentic AI has moved from slideware to budgets. Boards are asking for an agent strategy, vendors are shipping agents into every product, and most operations teams now have at least one agent pilot running.
Yet the results rarely match the activity. The organisations seeing measurable returns are not the ones with the most agents. They are the ones that fixed the process before they automated it.
The reason is simple: an agent executes a process faster, at scale, without getting tired. If the process is broken, the agent breaks it faster too. AI can’t outrun broken processes.
Adoption is rising. Returns are not.
Agent adoption is accelerating, but enterprise-level returns have stalled. McKinsey’s State of AI in 2026 survey, published in August 2026, makes the gap plain:
- Agents are scaling. 40% of respondents from organisations with more than $1 billion in revenue report scaling AI agents, up from 27% a year earlier.
- Profit isn’t following. Only 37% attribute any EBIT impact to AI, essentially unchanged from 2025.
- Few see real value. Only about 6% qualify as high performers, attributing at least 5% of EBIT to AI.
Gartner expects much of today’s agent activity to be written off. It predicts that over 40% of agentic AI projects will be cancelled by the end of 2027. The reasons it gives are escalating costs, unclear business value and inadequate risk controls.
What separates the 6% is not the model they use. Nearly three-quarters of McKinsey’s high performers have fundamentally redesigned workflows around AI, compared with a quarter of everyone else. They are also twice as likely to have defined processes for measuring impact.
Why agents amplify broken processes
An agent does not know a process is broken. It follows the rules, data and handoffs it is given, so every flaw in them gets executed faster and more often.
Four weaknesses do most of the damage:
- Data that disagrees with itself. When stock levels differ between the ERP, the warehouse spreadsheet and the CRM, an agent picks one. If it picks the wrong one, it promises a delivery date nobody can meet.
- Exceptions handled in email. Most processes work for the 80% of routine cases. The other 20% get resolved in inboxes and hallway conversations, which an agent cannot see. So it either stalls or guesses.
- No clear owner. If nobody owns the end-to-end process, nobody owns the agent’s output either. Errors bounce between teams instead of being fixed at the source.
- Approval steps that exist for historical reasons. Automating a five-step approval chain makes a slow process slightly faster. Removing the three steps that add no control makes it fundamentally faster.
This is why so many agent projects show promise in a pilot and then disappoint in production. A pilot runs on clean, curated data with a motivated team watching every step. Production runs on the real process, with all its workarounds.
What the organisations getting results do differently
Organisations that turn agents into measurable outcomes treat agentic AI as a process change programme with a technology component, not the other way round. Five habits stand out.
- They start with a business number, not a use case. The brief is “cut late customer orders by a third”, not “deploy an inventory agent”. The agent is whatever gets the number moving.
- They redesign the workflow before they automate it. They map how the work really happens, remove steps that add no control and decide which decisions belong to people and which to agents.
- They give agents one source of truth. Customer, order, stock, production and finance records sit on a single data model, so the agent never has to reconcile conflicting versions of reality.
- They define where humans stay in the loop. Agents recommend and prepare. People approve consequential actions such as purchase orders, price changes and schedule changes, and every approval is recorded.
- They measure from day one. A baseline is agreed before the agent goes live, and the agent is judged on that business number, not on how many tasks it completed.
None of these habits is technical. All of them are what make the technology pay back.
Fix the process before the agent
Every core operational process has a predictable weak point that will undermine an agent if it is left in place. Fix that first, then automate.
| Process | Where it usually breaks | Fix before deploying an agent | What the agent can then do |
| Order to cash | Sales promises dates without seeing real stock or capacity | Put orders, allocations and production on the same records | Answer “can we ship by Friday?” in seconds and flag at-risk orders |
| Procure to pay | Purchase orders raised by email; supplier dates never updated | Make the PO and confirmed delivery date the single source | Spot late suppliers and show which customer orders they affect |
| Plan to produce | Production schedule lives in a spreadsheet | Schedule work orders against live materials and capacity | Check clear-to-build before release and propose resequencing |
| Inventory and warehouse | Stock counted in one system, picked from another | One inventory ledger with lot and location detail | Summarise shortages, expiry risk and reorder needs |
| Quality and traceability | Batch records on paper or in separate files | Lot and serial genealogy recorded at each step | Answer which batches went to which customers in minutes |
| Field service | Engineers and parts tracked separately from customer history | Service, parts and installed base on the customer record | Prepare the right parts and history before the visit |
The pattern is the same in every row. The agent becomes useful only once the data it needs sits in one place and the handoffs are clear.
Measure agents in business terms
Agents should be measured the way the business is measured: in cash, time, service and risk. Counting tasks completed or hours “saved” says nothing about whether the P&L moved.
A practical scorecard tracks one leading signal and one business outcome per agent:
| Agent purpose | Leading signal (weekly) | Business outcome (quarterly) |
| Order promising | Share of recommendations accepted by sales | On-time delivery, order cancellations |
| Shortage and supplier risk | Shortages flagged before the work order is released | Expediting spend, production stoppages |
| Production readiness | Work orders released clear to build | Schedule changes, work in progress paused |
| Inventory health | Expiry and excess alerts actioned | Write-offs, inventory turns |
| Traceability | Time to answer a batch or recall query | Audit findings, recall cost |
| Service preparation | First-visit parts availability | First-time fix rate, repeat visits |
Two rules keep the scorecard honest:
- Agree the baseline before go-live. Without it, any improvement is an opinion.
- Stop agents that don’t move the number. If the business outcome has not shifted after a quarter, change the process or retire the agent. Don’t extend the pilot.
How Axolt approaches agentic AI
Axolt is built on the view that agents are only as good as the operations underneath them. That is why Axolt starts with the process and the data, then adds the agent.
Inventory, manufacturing, procurement, logistics, finance and field service all run natively inside Salesforce, on the same data model as your customers, opportunities and orders. There is no separate ERP to reconcile and no integration layer for an agent to fall through.
Axo AI, built on Agentforce, works directly on those live operational records. It can help teams:
- move from stock questions to coordinated action, through Agentforce inventory operations
- confirm production is clear to build before work orders are released
- surface shortages, supplier delay impact and expiry risk
- answer traceability questions from batch and lot records
Because it runs inside Salesforce, Axo AI works where your people already work, on the records they already trust.
Fix the process, then let the agent run
The organisations winning with agentic AI didn’t find a better model. They fixed the process, connected the data and decided in advance which number the agent had to move.
AI can’t outrun broken processes. But on a clean, connected process, it can run further and faster than any team could alone.
Bring one process that costs you time or money today, and we’ll show you what it looks like fixed and automated in Axolt. Book a demo
About Axolt
Axolt is a Salesforce-native ERP platform that brings Customer 360 and Operations 360 together on one platform. Manufacturers, distributors, medical device and pharmaceutical companies, and field service organisations use Axolt to run their operations inside Salesforce, on the same records as their customers and orders.
Axolt covers inventory and warehouse management, manufacturing ERP, MRP, production planning and scheduling, MES, quality control, procurement, logistics and multi-carrier shipping, finance, CPQ, field service, and booking and scheduling. Axo AI, built on Agentforce, brings agentic AI directly into those workflows.
Learn more at axolt.com.
Sources
- McKinsey & Company, The state of AI in 2026: On the road to ROI, August 2026
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 2025
