Search

What are you looking for?

Search our services, use cases and practical insights.

Enter at least 2 characters

Popular starting points

AI decisions

How do you choose an AI platform?

A good platform selection does not begin with the longest feature list. It starts with the applications, data flows and responsibilities the organisation needs to support consistently over time.

The short answer

Choose an AI platform against prioritised use cases and binding requirements for models, data, identity, integration, quality, security, operations, cost and portability. Test the shortlist with internal cases and a real integration before making a broad commitment.

In brief

  • A platform is useful when several applications genuinely need shared foundations.
  • Must-have requirements for data, access and operations should be fixed before product demonstrations.
  • A large model catalogue matters less than measurable quality and practical portability.
  • A pilot must prove identity, data flow, monitoring and exit, not merely generate an answer.

Which shared purpose should the platform serve?

A platform brings together reusable capabilities such as model access, identity, data integration, tool management, evaluation and monitoring. First identify which concrete applications need these capabilities during the next twelve to twenty-four months.

One use case does not automatically justify a broad platform. Conversely, several isolated projects create unnecessary duplication when each recreates the same access, data connections and controls.

  • Prioritised applications and user groups
  • Shared models, data sources and tools
  • Binding identity, quality and governance rules
  • Expected teams, volume, regions and operating hours

Which requirements belong in the comparison?

Separate mandatory criteria from desirable features. An exclusion concerning data location or identity integration must not be offset by many less important functions. Express requirements as testable scenarios rather than broad claims.

Which requirements belong in the comparison?
Mandatory gateTestable scenarioPass condition
Data and contractTrace the exact service data flow including logs, support and subprocessorsRegion, retention, training, deletion and contract meet the binding requirement
Identity and permissionsConnect SSO and test inherited document access with two different rolesThe platform never exposes more data than the source system permits
Quality and securityRun internal normal, edge and attack cases repeatedlyEvery minimum threshold is met and no defined critical failure occurs
OperationsSimulate a model change, quota, outage, alert and support escalationOwners receive actionable logs and the fallback path works
ExitMove configuration, test cases, data and application logic to an alternativeThe export is technically usable and contractually free of unclear residual dependency

What must a selection pilot prove in practice?

A prepared demo rarely reveals difficult boundaries. Ask candidates to implement the same representative use case with your languages, permissions and data conditions. The pilot should include failure cases and operational tasks.

  1. Step 1

    Quality

    Run internal test cases repeatedly and assess critical failures separately.

  2. Step 2

    Integration

    Connect single sign-on, roles, one real data source and one controlled tool call.

  3. Step 3

    Operations

    Test logs, cost, model versions, outages and support paths in practice.

  4. Step 4

    Exit

    Export or transfer data, configuration, test cases and application code to an alternative.

What does a weighted decision between two candidates look like?

Once mandatory gates have passed, remaining differences are rated on a fixed scale from 1 to 5. This anonymised scorecard deliberately shows why a high total cannot override an exclusion. Every score points to a test, contract term or operational record.

What does a weighted decision between two candidates look like?
CriterionWeightCandidate ACandidate BEvidence
Quality on internal cases30 %4/55/5Repeatable evaluation with the identical test set
Data and security25 %5/54/5Data flow, contract and technical security test
Integration and permissions20 %3/54/5, but permission inheritance is missingSSO, role and document-access test
Operations and support15 %4/53/5Outage exercise, logs and support case
Fully loaded cost and exit10 %2/54/5Three-year cost model and export test
Weighted result100 %77/100; all gates passed83/100; excluded because permission inheritance is missingRecorded decision with accepted trade-offs

How does selection become a sustainable decision?

Business, architecture, security, privacy, procurement and operations should assess the same documented criteria from their respective responsibilities. The decision records not only the winner but assumptions, open risks and accepted dependencies.

  • Freeze weighting and exclusion criteria before final scoring
  • Map contractual claims to the exact product, plan and operating mode
  • Name owners and budgets for the platform and initial applications
  • Reassess after the pilot, a contractual change or major architecture change
Example from day-to-day business

Example: one platform for three initial applications

A company plans knowledge search, document processing and a service copilot. Two platforms meet the quality requirements. The pilot shows that only one inherits existing permissions down to document level, sends logs to central monitoring and supports model changes with the same evals. Export of configuration and data is tested contractually and technically before broader rollout.

What to remember

Choose a platform for a clear shared purpose. Prove mandatory criteria with internal cases and test data flow, operations and exit as practically as the visible AI capabilities.

Sources and further reading

These primary sources provide further detail on definitions, technical foundations or responsible use.

Content reviewed

Would you like to apply this to your situation?

Together, we clarify what makes sense for your process, data and systems – in plain language and without unnecessary complexity.

Discuss Your Project