Fractal connects enterprise AI agents with business processes and data

Fractal combines AI engineering, business-process expertise and its Cogentiq agent platform to help enterprises put AI into everyday operations. The practical buying question is whether a process needs a reusable platform, a tailored implementation, or both. A useful evaluation follows one decision from source data through human approval into the system where work is recorded.

01 / ProductHow Cogentiq fits Fractal’s broader AI business

The Fractal site organizes its work around AI-led transformation, AI foundations, and work and workforce adoption. That is a broader engagement than purchasing a chatbot. A project may need data integration, evaluation methods, redesigned responsibilities and training alongside an AI interface. Fractal positions Cogentiq as a common platform for that work, while its product portfolio also includes function-specific offers for commerce, customer experience and invoice-to-cash operations.

Cogentiq describes connectors for databases, data lakes, documents and APIs; agent orchestration; low-code and code-based building; and human review controls. Its published feature set includes stateful execution and recovery, model choice, access controls and audit trails. These are capabilities to examine in a configured environment, not evidence that every connected workflow is ready to act safely. A connector can expose a record without resolving what that record means to a finance team.

Fractal’s history also includes investment in Qure.ai. Its FY2023 financial statements record loss of control in 2022 and associate-company treatment. This blueprint covers Fractal’s enterprise workflow offer. Qure.ai’s medical-imaging products have a distinct clinical purpose and their own coverage; a Fractal agent engagement should not be assumed to include them.

02 / AudienceWho benefits from a platform plus implementation partner

A strong candidate has a repeated operational bottleneck whose evidence is split across systems. Collections staff may have invoices in an ERP, delivery evidence in document storage and customer correspondence in a shared mailbox. They need a coherent worklist and a defensible explanation of each suggested action. Fractal becomes relevant when connecting these elements and changing the process are substantial parts of the job.

A business sponsor and a technical owner should work together. The sponsor decides which exceptions are costly and what authority staff may delegate. The technical owner maps credentials, interfaces, record identifiers and failure recovery. Subject-matter experts then supply examples of valid and invalid decisions. Without these roles, a polished demonstration can hide unresolved questions about who is allowed to change a customer balance or promise a payment arrangement.

A small team seeking only personal writing assistance probably has less need for this engagement model. Likewise, an organization that already operates an effective agent stack should identify a specific missing capability before introducing another platform. Dataiku is a useful adjacent comparison for teams evaluating a shared environment for data and AI development; Fractal’s central question is how much implementation and process redesign the buyer wants delivered with the technology.

03 / WorkflowA proposed invoice-dispute workflow

Consider a distributor trying to resolve deductions on customer payments. Cogentiq Invoice to Cash describes invoice validation, collections worklists, risk profiling and deduction-resolution workflows. The following is an illustrative implementation plan, not a claim that this exact workflow is available without configuration.

1. 01

   Define the disputed amount

   Start with a closed historical sample: invoice, payment, deduction reason, shipment evidence and final resolution. Preserve both the original customer identifier and the ledger identifier so similarly named accounts cannot be merged by the assistant.

2. 02

   Build a bounded evidence packet

   Let a read-only agent assemble relevant documents and correspondence for a reviewer. Require dates and record references with each conclusion. A missing delivery receipt should remain a missing receipt; it should not become an assumed proof of delivery.

3. 03

   Separate explanation from accounting

   Use AI to summarize the customer’s position and suggest the next question. Calculate balances and due dates through approved business logic. A language model should not independently reinterpret the ledger or create a credit merely because an email sounds persuasive.

4. 04

   Review before taking action

   Have a collections specialist approve the proposed response and any recommended accounting adjustment. Record the reviewer, the evidence version and the exact permitted action. Escalate contradictory records, disputed contractual terms and unusually large deductions.

5. 05

   Reconcile the result

   After an approved update, read back the destination system and attach the resulting reference to the case. Measure unresolved exceptions, reviewer corrections and repeat disputes alongside handling time. An attempted action is not the same as a reconciled balance.

Fractal’s enterprise-context whitepaper argues for representing business entities, relationships and rules explicitly. In this example, that means defining the relationship between an invoice, a shipment, a deduction and an authorized approver. The useful test is whether those relationships remain intact when evidence is incomplete, rather than whether the agent can produce a fluent paragraph.

04 / PricingCommercial scope should follow the process boundary

The consulted platform and product pages direct buyers toward a demo or consultation; they do not provide a generally applicable public subscription tariff. A Fractal consultation therefore needs a defined scope. Do not price an invoice-to-cash project by multiplying a guessed per-seat charge by the collections team’s headcount.

Separate initial integration from ongoing operation in the proposal. A process can be inexpensive to demonstrate using exported files yet costly to maintain when upstream schemas, access policies or document templates change. Ask for the acceptance criteria for each connector and a clear owner for failures. Also distinguish charges paid to Fractal from any cloud or model-provider charges paid directly by the customer.

05 / DistinctionsEnterprise context and adoption are part of the work

The meaningful distinction is the combination of business-specific agents with engineering and adoption services. Fractal’s workforce offering explicitly addresses applying AI in work and measuring adoption. For the proposed collections project, adoption means whether staff trust and use the evidence packet, whether corrections are incorporated, and whether the tool reduces unresolved exceptions. Attendance at training alone would not answer those questions.

Accenture AI Refinery offers another reference point for coordinated enterprise agents and human interaction. Compare the concrete handoff between platform, integrator and internal operators. A buyer that wants a tailored outcome may value implementation depth; a buyer building a durable internal capability may place greater weight on editable workflows, documentation and the ability to change providers later. Neither preference establishes a universal winner.

06 / QuestionsQuestions to settle before agents touch live records

Can the workflow recover without repeating a financial action?

Ask the team to demonstrate a timeout after an approved action has reached the ERP but before the agent receives confirmation. The correct recovery should establish whether the update already happened. Replaying a credit or reminder blindly can create duplicate work even when every individual model response is sensible. Inspect the application-level evidence rather than treating a stateful-execution feature as proof of business-level reconciliation.

What does model or cloud independence mean for this project?

Cogentiq advertises a cloud- and model-agnostic architecture. The buyer still needs to know which components can actually move: prompts, agent definitions, stored context, retrieval indexes, credentials and evaluation sets. Request the supported deployment design and an export example for the proposed configuration. A platform-level portability claim does not settle the cost of replacing a customized integration.

Who approves the business definitions?

A shared context layer needs accountable owners. Finance may define an overdue invoice differently from sales, and a dispute can suspend collection without cancelling the amount owed. Keep those distinctions explicit and versioned. This public-source research does not verify Fractal’s accuracy, implementation speed or customer returns; those must be assessed against the buyer’s own cases and operating constraints.

07 / DecisionChoose Fractal around an accountable operational outcome