- Integration, not model quality, is the blocker. 56% of enterprise leaders in a 2026 Gartner survey named integrating AI with legacy systems and processes as their main challenge, ahead of talent at 50%.
- Sequencing AI behind a platform rebuild is what kills the business case. The foundation is never finished, and the roadmap inherits an ERP timeline.
- An AI orchestration layer sits above the stack, connects by API rather than by screen, and needs no data migration.
- Four design choices decide whether it survives audit: API not interface, inherited permissions, live reads, and approved writes with an audit trail.
- A first production deployment is a quarter, not a programme. One decision, two or three systems, one write-back action.
Part 1 · The rebuild reflex
The advice you keep getting is to fix the foundation first
Consolidate, modernise, clean, then add intelligence. It sounds prudent. It is how AI roadmaps get pushed to 2029.
Ask most advisors how to get AI working across an enterprise and you get a sequencing answer. Consolidate the data. Modernise the ERP. Stand up the warehouse. Clean the master data. Then, once the foundation is right, put intelligence on top of it.
There is real logic in that. There is also a reason so many AI programmes now carry a three-year runway and a business case that nobody can defend by month eighteen. The foundation is never finished. Every enterprise that has completed a data platform project has started the next one, and the AI use cases waiting politely behind it are still waiting.
Meanwhile the systems themselves are usually fine. The ERP posts what it was bought to post, the CRM holds the commitments, the service desk tracks the tickets. None was built to answer a question crossing all three, but that is a gap between them, not a defect inside them.
What is an AI orchestration layer?
Part 2 · The evidence
What the data says is actually blocking you
Enterprises are not short of AI. They are short of AI that reaches the systems where work is recorded.
Gartner surveyed 140 senior supply chain leaders at organisations above $250 million in revenue in late 2025. The single most cited challenge in scaling AI, at 56%, was integrating it with legacy systems and processes. Talent came second at 50%. Gartner's own framing is blunt: the greatest friction is not the technology, but the legacy environments in which it is being deployed.
Put that next to the number everyone quoted through 2025. MIT researchers reported that roughly 95% of enterprise generative AI pilots produced no measurable return. The detail worth more than the headline is that vendor-led and partner-built deployments succeeded around twice as often as tools built in house, and the ones that worked plugged into how people already worked instead of asking them to work somewhere new.
McKinsey's 2025 survey found 88% of organisations using AI in at least one function, and 7% saying it was fully scaled. Gartner separately expects more than 40% of agentic AI projects to be cancelled by the end of 2027, on escalating cost, unclear business value and weak risk controls, noting that pushing agents into legacy systems tends to require costly modifications.
Read those four numbers together and a pattern shows up that has nothing to do with model capability. Adoption is close to universal. Scaling is rare. The step that fails is the last mile into the systems of record.
Gartner makes the serious counter-argument: bolting AI onto an analog-era foundation can lock in the inefficiency you already had. Fair, where the process itself is broken. That is a case for fixing named processes, not for holding every use case hostage to a migration.
Eerly AI StudioPart 3 · The overlay
Put the intelligence above the stack, not inside it
Vendor AI lives inside one application. RPA lives on top of one screen. A layer lives above all of them.
Deploying enterprise AI without a rebuild is a placement decision before it is a product decision. Instead of embedding intelligence inside each application, or scripting a bot per process, you run one horizontal layer above the estate that reaches into SAP, Oracle, Workday, ServiceNow, Salesforce and the document stores through their APIs.
Four design choices separate a layer that survives an audit from one that dies in pilot.
Robotic process automation drives the screen, which is why a UI change breaks the bot and every new process needs a new bot. API-based integration with reasoning between the steps handles the exception instead of failing on it, and one platform covers many processes rather than one script each.
Read entitlements from each source system at query time. A user should see exactly what their role already allowed and nothing more. The moment a deployment copies enterprise data into a separate index with its own weaker access model, you have built a governance problem that will outlive the pilot.
An answer built on last night's export is a confident answer about yesterday. Reading through to the source keeps the response current and removes the second copy of your data that nobody wanted to maintain.
Answering alone leaves the manual entry, and its error rate, exactly where it was. The layer should execute the approved change in the system of record and return evidence that it completed, with a person approving by default and every action logged.
Part 4 · The first quarter
What a first ninety days actually looks like
One decision. Two or three systems. One action that writes back.
The programmes that stall tend to start wide: connect everything, index everything, then find a use case. The ones that land start with a decision somebody makes every week and cares about.
Pick one recurring cross-system decision and time it honestly. Log the hour it was first asked and the hour someone acted. That interval is your baseline.
Connect the two or three systems holding the pieces, read-only, inheriting each system's permission model.
Put grounded answers in front of one team, with the source system and timestamp on every fact. People check AI against systems they already trust, so give them the receipts.
Turn on one write-back action behind human approval. A master data update or a requisition is enough to close the loop.
Notice what is absent. No warehouse, no re-platforming, no parallel copy of the ERP. The scope is deliberately narrow, because a narrow deployment produces a number, and a number is what buys you the second use case.
Part 5 · Non-negotiables
Three things not to trade away for speed
Deploying fast is easy. Deploying fast in a way security will sign off twice is the actual skill.
Deploy AI across the systems you already run.
One AI layer above SAP, Oracle, Workday and ServiceNow. It answers inside the permissions you already enforce, with a full audit trail. No rip and replace.
Book a demo →Part 6 · The takeaway
You already bought the hard part
The expensive work is done. What is missing is a layer that can read across it.
Your ERP took years and a small fortune. Your CRM holds a decade of customer history. Your service desk encodes how the organisation behaves under pressure. That is the hard part, and it is finished.
What is missing is not another platform underneath it. It is something above it that reads across all of it, inside the permissions you already enforce, and puts a defensible answer in front of whoever has to decide. The enterprises pulling ahead on AI in 2026 did not rebuild first. They stopped waiting for a foundation that was never going to be finished, and started with one decision, three systems and ninety days.
FAQ
Frequently asked questions
What enterprise teams ask us most about deploying AI on existing systems.
Can you deploy AI without replacing your ERP or legacy systems?
Yes. An AI orchestration layer connects to ERP, CRM, HR and ITSM systems through their existing APIs, reads live data within each user's current permissions, and writes approved actions back into the source system. Nothing is migrated and nothing is replaced. The systems keep operating exactly as they do today, and the intelligence sits above them rather than inside any one of them.
Why do most enterprise AI pilots fail to scale?
Not because the models are weak. Gartner's 2026 supply chain survey found 56% of leaders name integrating AI with legacy systems and processes as their main challenge, ahead of talent shortages at 50%. MIT research reported that around 95% of enterprise generative AI pilots produced no measurable return, with vendor-led deployments succeeding roughly twice as often as tools built in house. The pattern points to the last mile into systems of record, not to model quality.
Should we modernise our data platform before deploying AI?
Only where a specific process is genuinely broken. Sequencing every AI use case behind a platform migration ties the AI roadmap to an ERP timeline, and the foundation is never finished. A better test is per use case: if the data needed for a decision exists somewhere and is readable through an API, that use case does not need to wait for the migration.
How is an AI orchestration layer different from RPA?
RPA scripts the user interface, so a screen change breaks the bot and each new process needs a new bot. An orchestration layer works through APIs with reasoning between the steps, which means it can handle exceptions rather than failing on them, and one platform covers many processes instead of one script per process.
Does adding an AI layer create a new security or access risk?
It should not, provided the layer inherits access control instead of declaring its own. Permissions are read from each source system at query time, so a user sees only what their role already allowed. Risk enters when a deployment copies enterprise data into a separate index governed by a weaker, parallel permission model. Deploying inside the organisation's own cloud, VPC or data centre keeps data inside the existing perimeter.
How long does it take to deploy AI across existing enterprise systems?
A focused first deployment is a quarter of work rather than a multi-year programme: roughly two weeks to pick and baseline one recurring decision, two weeks to connect two or three systems read-only, a month to put grounded answers with provenance in front of one team, and a final month to enable a single write-back action behind human approval.
What does it mean for AI to act in the source system?
It means the AI does not stop at telling someone what to do. After a human approves, it performs the update in the system of record itself, such as a master data change or a purchase requisition, and returns evidence that it completed along with a logged audit trail. Answering alone leaves the manual entry, and the error rate that comes with it, exactly where it was.
Sources & further reading
- Gartner. Survey Finds Technology Integration and Talent Perceived as Key Roadblocks to Scaling AI in Supply Chain (April 2026). Survey of 140 senior supply chain leaders at organisations with revenue of $250M or more, conducted October to November 2025: 56% cite integrating AI with legacy systems and processes, 50% cite limited internal expertise. gartner.com
- Gartner. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 2025). Cancellation drivers: escalating costs, unclear business value, inadequate risk controls; integration into legacy systems noted as technically complex and often requiring costly modifications. gartner.com
- MIT Media Lab, Project NANDA. The GenAI Divide: State of AI in Business 2025 (August 2025), as reported. Approximately 95% of enterprise generative AI pilots delivered no measurable ROI; vendor-led deployments outperformed internal builds. healthcareitnews.com
- McKinsey & Company. AI at Work, but Not at Scale, drawing on the State of AI survey (2025). 88% of organisations report using AI in at least one business function, up from 78% in 2024; 7% report AI fully scaled. mckinsey.com
AI expert driving the design and architecture of innovative AI solutions. Seamlessly integrating cutting-edge models into scalable systems, leads with a focus on transforming complex data into actionable insights, delivering exceptional business outcomes.