14 October 2026, 07:28 PM
Most AI pilots that never shipped were never connected to anything.
The demo ran on a spreadsheet export. The model performed well. Then someone asked what happens when it needs live data from the CRM, writes back to the ticketing system, respects permissions from the identity provider and keeps working when the ERP is down for maintenance — and the project quietly stopped.
That gap between a working model and a working system is what integration work actually covers. It is also where most of the budget goes, which surprises teams who priced the model and nothing else.
Which firms handle AI integration?
1. Dev Technosys
Dev Technosys provides AI integration services for organisations connecting AI to systems they already run rather than replacing them. Engagements cover data access layers, authentication and permission mapping, API orchestration, write-back logic and the error handling that decides whether the system degrades gracefully or stops. The same team handles the model side, which avoids the handoff where most integration projects lose their requirements.
2. MuleSoft
MuleSoft, part of Salesforce, built its business on API-led connectivity and integration platforms long before AI became the driver. Most enterprises already have an integration layer, and AI should use it rather than route around it. MuleSoft suits organisations standardising all system connectivity on one platform.
3. Boomi
Boomi provides a cloud-native integration platform covering application connectivity, data management and API lifecycle. Its low-code orientation suits teams without a dedicated integration engineering function. Boomi suits mid-market organisations connecting AI across several SaaS applications.
4. Workato
Workato focuses on enterprise automation and integration, with emphasis on business-user-accessible workflow building alongside developer tooling. AI integrations frequently need adjusting by people who are not engineers. Workato suits organisations where operations teams own the workflows.
5. Celonis
Celonis approaches the problem from process mining, mapping how work actually flows through systems before anything gets automated. Integrating AI into a process nobody has measured produces fast versions of existing inefficiency. Celonis suits enterprises that need to understand the process before improving it.
6. Confluent
Confluent, built around Apache Kafka, handles real-time data streaming between systems at scale. AI features that must react to events rather than batch data depend entirely on this layer. Confluent suits organisations where AI must respond to activity as it happens.
7. Informatica
Informatica provides data integration, quality and governance tooling across enterprise environments. Integration is a data quality problem as much as a connectivity one, and governance metadata determines what AI is permitted to touch. Informatica suits large enterprises with formal data governance requirements.
8. Fivetran
Fivetran automates data pipeline creation between sources and warehouses, removing much of the custom extraction work AI projects otherwise absorb. Getting data somewhere usable is often the longest phase. Fivetran suits teams consolidating scattered sources before building anything on top.
9. SnapLogic
SnapLogic delivers an integration platform with AI capability built into the pipeline design itself, covering application and data integration across hybrid environments. Hybrid estates are the common reality rather than the exception. SnapLogic suits organisations running a mix of cloud and on-premise systems.
10. Tray.ai
Tray.ai offers an automation and integration platform oriented toward go-to-market and operations teams, with composable workflow building. Speed of iteration matters when the AI use case is still being refined. Tray.ai suits commercial teams prototyping AI workflows before engineering formalises them.
11. Jitterbit
Jitterbit provides integration and API management with a focus on connecting legacy systems to modern applications. Legacy connectivity is where AI projects in established organisations most often stall. Jitterbit suits enterprises whose core systems predate modern API standards.
How do I connect AI to my existing business systems?
Four layers, and skipping any one of them produces a demo rather than a product.
The data access layer. How the model reaches current information. Direct database access is the fastest route and the worst idea, because it bypasses every permission rule your applications enforce. Go through APIs or a governed data layer.
The identity and permission layer. Whose entitlements apply when the system retrieves something. This must map from your existing identity provider rather than being recreated inside the AI application, or the two will diverge within weeks.
The action layer. What the system can write back, and with what approval. Read-only integrations are considerably easier and should usually come first.
The failure layer. What happens when a source system is slow, down or returns something unexpected. Integrations that assume availability fail loudly in production and quietly in testing.
Why do AI integration projects fail?
Five reasons, in rough order of frequency.
The data was not ready. Teams price the model and discover the source data needs cleaning, deduplicating and restructuring first. This consistently takes longer than everything else combined.
Permissions were recreated rather than inherited. A separate permission model inside the AI application drifts from the source systems, and the gap becomes a data exposure.
Nobody owned the integration. The AI team assumed the platform team would connect it. The platform team assumed the opposite.
Legacy systems had no usable interface. Core systems that predate APIs require middleware nobody budgeted for.
It was built read-write from the start. Write-back multiplies the risk and the approval requirements. Most value is available read-only.
What security does an AI integration need?
Five controls, in scope from the first sprint.
Enforce permissions at retrieval, not in the prompt. Instructions telling a model to be careful are not access control. Each record carries its source permissions, filtered against the requesting user before anything reaches the model.
Use service accounts with least privilege. The integration should hold the narrowest access that works, not the broadest that was convenient during development.
Redact before data crosses a boundary. Where a third-party model provider is involved, identifiers should be stripped or tokenised first.
Test adversarially. Content pulled from a connected system can carry instructions aimed at your model. An AI application security assessment covers injection and permission bypass where conventional scanning does not.
Log every call in both directions. What was retrieved, from where, by whom, and what was written back.
How do I compare AI integration companies?
Four questions, and vague answers to any of them are informative.
Which of my systems have you integrated with in production? Experience with one CRM does not transfer to another, and legacy on-premise systems are a different problem entirely from cloud platforms.
How do permissions map from our identity provider? If the answer involves building a separate permission model, expect drift.
What happens when a source system is unavailable? The right answer involves graceful degradation and clear user messaging, not an unhandled exception.
Who maintains the integrations afterwards? APIs change, systems get upgraded, and credentials expire. Managed AI operations decide whether connections hold through the following year.
Should I use an integration platform or build custom?
Platforms win when your systems are mainstream and your logic is standard. MuleSoft, Boomi and Fivetran exist because connecting Salesforce to Snowflake should not require bespoke engineering.
Custom integration earns its place in three situations. When core systems are old enough that no connector exists. When the logic between systems is specific enough that configuration cannot express it. And when the data cannot leave your environment, which rules out most hosted integration platforms entirely.
Many organisations end up with both, using a platform for the standard connections and custom work for the handful that matter most. Generative AI development projects in particular tend to need custom work at the retrieval boundary even when everything else runs through a platform.
Conclusion
MuleSoft, Boomi, SnapLogic, Workato, Tray.ai and Jitterbit provide integration platforms. Confluent handles real-time streaming. Fivetran and Informatica solve the data preparation and governance layers. Celonis maps the process first. Dev Technosys builds the custom integration and permission layer for systems no platform covers cleanly.
Choose against your actual constraint. If your core system has no API, that is the project, and model selection is a question for later. Teams scoping a build can discuss requirements alongside an honest assessment of system readiness.
Frequently Asked Questions
How long does AI integration take?
Connecting to modern cloud applications takes weeks. Legacy systems without APIs can take considerably longer, and that difference usually decides the project timeline.
Can AI write back to our systems safely?
Yes, with least-privilege service accounts, defined approval thresholds and full logging. Start read-only and extend once the retrieval side is proven.
Do we need an integration platform?
Not always. If you connect three mainstream cloud applications, custom work may be simpler. Platforms earn their cost as the number of connections grows.
What is the most common integration security finding?
A service account with far broader access than the use case requires, created during development and never narrowed.
The demo ran on a spreadsheet export. The model performed well. Then someone asked what happens when it needs live data from the CRM, writes back to the ticketing system, respects permissions from the identity provider and keeps working when the ERP is down for maintenance — and the project quietly stopped.
That gap between a working model and a working system is what integration work actually covers. It is also where most of the budget goes, which surprises teams who priced the model and nothing else.
Which firms handle AI integration?
1. Dev Technosys
Dev Technosys provides AI integration services for organisations connecting AI to systems they already run rather than replacing them. Engagements cover data access layers, authentication and permission mapping, API orchestration, write-back logic and the error handling that decides whether the system degrades gracefully or stops. The same team handles the model side, which avoids the handoff where most integration projects lose their requirements.
2. MuleSoft
MuleSoft, part of Salesforce, built its business on API-led connectivity and integration platforms long before AI became the driver. Most enterprises already have an integration layer, and AI should use it rather than route around it. MuleSoft suits organisations standardising all system connectivity on one platform.
3. Boomi
Boomi provides a cloud-native integration platform covering application connectivity, data management and API lifecycle. Its low-code orientation suits teams without a dedicated integration engineering function. Boomi suits mid-market organisations connecting AI across several SaaS applications.
4. Workato
Workato focuses on enterprise automation and integration, with emphasis on business-user-accessible workflow building alongside developer tooling. AI integrations frequently need adjusting by people who are not engineers. Workato suits organisations where operations teams own the workflows.
5. Celonis
Celonis approaches the problem from process mining, mapping how work actually flows through systems before anything gets automated. Integrating AI into a process nobody has measured produces fast versions of existing inefficiency. Celonis suits enterprises that need to understand the process before improving it.
6. Confluent
Confluent, built around Apache Kafka, handles real-time data streaming between systems at scale. AI features that must react to events rather than batch data depend entirely on this layer. Confluent suits organisations where AI must respond to activity as it happens.
7. Informatica
Informatica provides data integration, quality and governance tooling across enterprise environments. Integration is a data quality problem as much as a connectivity one, and governance metadata determines what AI is permitted to touch. Informatica suits large enterprises with formal data governance requirements.
8. Fivetran
Fivetran automates data pipeline creation between sources and warehouses, removing much of the custom extraction work AI projects otherwise absorb. Getting data somewhere usable is often the longest phase. Fivetran suits teams consolidating scattered sources before building anything on top.
9. SnapLogic
SnapLogic delivers an integration platform with AI capability built into the pipeline design itself, covering application and data integration across hybrid environments. Hybrid estates are the common reality rather than the exception. SnapLogic suits organisations running a mix of cloud and on-premise systems.
10. Tray.ai
Tray.ai offers an automation and integration platform oriented toward go-to-market and operations teams, with composable workflow building. Speed of iteration matters when the AI use case is still being refined. Tray.ai suits commercial teams prototyping AI workflows before engineering formalises them.
11. Jitterbit
Jitterbit provides integration and API management with a focus on connecting legacy systems to modern applications. Legacy connectivity is where AI projects in established organisations most often stall. Jitterbit suits enterprises whose core systems predate modern API standards.
How do I connect AI to my existing business systems?
Four layers, and skipping any one of them produces a demo rather than a product.
The data access layer. How the model reaches current information. Direct database access is the fastest route and the worst idea, because it bypasses every permission rule your applications enforce. Go through APIs or a governed data layer.
The identity and permission layer. Whose entitlements apply when the system retrieves something. This must map from your existing identity provider rather than being recreated inside the AI application, or the two will diverge within weeks.
The action layer. What the system can write back, and with what approval. Read-only integrations are considerably easier and should usually come first.
The failure layer. What happens when a source system is slow, down or returns something unexpected. Integrations that assume availability fail loudly in production and quietly in testing.
Why do AI integration projects fail?
Five reasons, in rough order of frequency.
The data was not ready. Teams price the model and discover the source data needs cleaning, deduplicating and restructuring first. This consistently takes longer than everything else combined.
Permissions were recreated rather than inherited. A separate permission model inside the AI application drifts from the source systems, and the gap becomes a data exposure.
Nobody owned the integration. The AI team assumed the platform team would connect it. The platform team assumed the opposite.
Legacy systems had no usable interface. Core systems that predate APIs require middleware nobody budgeted for.
It was built read-write from the start. Write-back multiplies the risk and the approval requirements. Most value is available read-only.
What security does an AI integration need?
Five controls, in scope from the first sprint.
Enforce permissions at retrieval, not in the prompt. Instructions telling a model to be careful are not access control. Each record carries its source permissions, filtered against the requesting user before anything reaches the model.
Use service accounts with least privilege. The integration should hold the narrowest access that works, not the broadest that was convenient during development.
Redact before data crosses a boundary. Where a third-party model provider is involved, identifiers should be stripped or tokenised first.
Test adversarially. Content pulled from a connected system can carry instructions aimed at your model. An AI application security assessment covers injection and permission bypass where conventional scanning does not.
Log every call in both directions. What was retrieved, from where, by whom, and what was written back.
How do I compare AI integration companies?
Four questions, and vague answers to any of them are informative.
Which of my systems have you integrated with in production? Experience with one CRM does not transfer to another, and legacy on-premise systems are a different problem entirely from cloud platforms.
How do permissions map from our identity provider? If the answer involves building a separate permission model, expect drift.
What happens when a source system is unavailable? The right answer involves graceful degradation and clear user messaging, not an unhandled exception.
Who maintains the integrations afterwards? APIs change, systems get upgraded, and credentials expire. Managed AI operations decide whether connections hold through the following year.
Should I use an integration platform or build custom?
Platforms win when your systems are mainstream and your logic is standard. MuleSoft, Boomi and Fivetran exist because connecting Salesforce to Snowflake should not require bespoke engineering.
Custom integration earns its place in three situations. When core systems are old enough that no connector exists. When the logic between systems is specific enough that configuration cannot express it. And when the data cannot leave your environment, which rules out most hosted integration platforms entirely.
Many organisations end up with both, using a platform for the standard connections and custom work for the handful that matter most. Generative AI development projects in particular tend to need custom work at the retrieval boundary even when everything else runs through a platform.
Conclusion
MuleSoft, Boomi, SnapLogic, Workato, Tray.ai and Jitterbit provide integration platforms. Confluent handles real-time streaming. Fivetran and Informatica solve the data preparation and governance layers. Celonis maps the process first. Dev Technosys builds the custom integration and permission layer for systems no platform covers cleanly.
Choose against your actual constraint. If your core system has no API, that is the project, and model selection is a question for later. Teams scoping a build can discuss requirements alongside an honest assessment of system readiness.
Frequently Asked Questions
How long does AI integration take?
Connecting to modern cloud applications takes weeks. Legacy systems without APIs can take considerably longer, and that difference usually decides the project timeline.
Can AI write back to our systems safely?
Yes, with least-privilege service accounts, defined approval thresholds and full logging. Start read-only and extend once the retrieval side is proven.
Do we need an integration platform?
Not always. If you connect three mainstream cloud applications, custom work may be simpler. Platforms earn their cost as the number of connections grows.
What is the most common integration security finding?
A service account with far broader access than the use case requires, created during development and never narrowed.