AI systems are moving beyond drafts. They are choosing models, reading customer accounts, testing software, changing production code and turning live sports data into public content.

Each action carries a consequence. A poor model route wastes money. A weak support decision can breach regulation. A security agent can cross a legal boundary. A bad code change can break a product.

The five developments in this edition point to the same product problem. The model is only one part of the system. Access rules, routing logic, evaluation, human approval and clear evidence determine whether the system is fit for real work.

What this issue covers

Stripe’s proposed acquisition of OpenRouter and the economics of model routing
Fanatics’ use of specialist agents in regulated customer support
Digital Encode’s autonomous penetration-testing platform in Nigeria
Asana’s use of coding agents to remove long-standing technical debt
The USTA and IBM’s use of AI inside live editorial products

Stripe is moving closer to the AI decision layer

On 19 August, Stripe agreed to acquire OpenRouter, a gateway that lets developers access and route requests across more than 400 AI models from over 80 providers.

OpenRouter says it processes more than 10 trillion tokens a day for a community of over 10 million developers and companies. The transaction remains subject to closing conditions. OpenRouter says its name, product, roadmap and existing integrations remain unchanged for now.

Stripe and OpenRouter both describe the deal as an agreement, rather than a completed integration.

The applied workflow already exists inside OpenRouter. An application sends an inference request. The routing layer applies developer-set rules covering model choice, fallback, price, latency, provider availability and data retention. It sends the request to the selected provider and returns the response with usage and cost information.

Developers remain responsible for the routing policy.

Stripe already manages choices around payment methods, authorisation, fraud and billing. OpenRouter manages a different economic decision: which intelligence provider should handle a task at a given cost and level of reliability.

The agreement does not mean that AI agents can now pay autonomously or that the two products have been integrated.

For African product teams, the practical issue is less about accessing hundreds of models. It is about controlling cost, latency, availability and data treatment across a smaller set of useful providers.

Foreign-currency billing and uneven connectivity make failed or poorly routed requests expensive. The next evidence should show whether the combined business preserves provider neutrality, improves routing quality and gives customers clear control over where their data goes.

Fanatics gave different agents different responsibilities

Fanatics Betting and Gaming has put a multi-agent system into customer support. Customers ask about deposits, withdrawals, account settings, betting rules and responsible-gaming support.

The answers vary by US state, and support demand can rise sharply during live events.

The workflow begins with a customer message. A guardrail checks for prompt injection, while a classifier reviews responsible-gaming risk. A supervisor agent calls the appropriate knowledge, account or transaction tool.

High-severity cases transfer to a person with the conversation context.

This is a useful example of autonomy being divided by risk. The system can retrieve an account detail or answer a documented question. It cannot treat every interaction as an ordinary support ticket.

Fanatics started with four of more than 20 case types and kept explicit escalation paths for sensitive cases.

Fanatics reports that containment improved by about 56% and resolution by about 53% during the first two months. The system has handled thousands of cases, according to the joint Fanatics and AWS implementation account.

These are internal relative improvements. The publication does not provide the starting rates, customer-satisfaction figures, responsible-gaming error rates or independent validation.

The lesson transfers to regulated financial-services support. Product teams would still need country-specific rules, accurate account integrations, language testing and clear escalation for fraud, financial distress or customer vulnerability.

A support agent should earn wider scope one case type at a time.

Digital Encode is testing how far security agents should act

Nigeria’s Digital Encode launched DEPAS AI on 17 August. The platform uses specialist agents to test authorised web and mobile applications, APIs, backend systems, cloud environments and networks.

An agentic coordinator assigns work, a verification agent attempts to reproduce findings, and another component links weaknesses into possible attack paths.

At a Lagos demonstration, Digital Encode gave the system an intentionally vulnerable banking application and instructed it in ordinary English to begin testing.

The system identified weaknesses without requiring a security professional to operate every tool manually. It then attempted authorised exploitation and produced remediation guidance, according to Punch and BusinessDay.

The use case is relevant to Nigerian organisations that depend on connected digital systems. Continuous testing could find weaknesses sooner.

The target must be explicitly authorised, the permitted actions kept narrow and every attempted action recorded.

Digital Encode says the platform found more than 500 vulnerabilities for one Nigerian organisation. The customer is unnamed, and the claim has not been independently verified.

The company has not published false-positive rates, test coverage, approval gates or evidence that the reported findings were fixed. DEPAS AI should be treated as an early commercial deployment with an important workflow, rather than a proven replacement for skilled penetration testers.

Asana made an old engineering problem economically viable

Asana used coding agents to remove Enzyme, an outdated testing system that had blocked upgrades to its frontend codebase.

The previous staffing plan estimated at least five years of work and roughly $6 million in cost. Asana says the migration was completed in about two weeks, with model and infrastructure costs of around $12,000.

The mechanism matters more than the headline.

Up to four coding agents worked in parallel, each inside a separate copy of the codebase. An engineer checked progress twice a day and reviewed every proposed change before it entered the normal engineering process.

A short five-sentence prompt reportedly worked better than a more elaborate setup. OpenAI’s Asana case study supplies the time and cost figures, so they should be treated as customer- and vendor-reported rather than independently audited.

Technical debt often survives because the manual cost is difficult to justify against new features. Agents can change that calculation when the migration is bounded, tests exist and engineers retain merge authority.

The missing evidence is what happened after completion. Asana has not published defect rates, rollbacks, escaped regressions or the maintenance cost of the migrated code.

Teams considering similar work should compare agent-generated changes against the same quality gates used for human code and track problems after release.

The US Open is using AI inside an editorial product

IBM and the United States Tennis Association announced new AI features for the 2026 US Open on 24 August.

Live Updates personalises the home page around selected players. Key Moments explains swings in a match. Serve Quality analyses 21 points across a player’s body and racquet, tracked 50 times per second, to generate a near-real-time score.

IBM estimates that the tournament will produce about 1.2 billion data points for this feature.

The workflow connects match data, expert opinion and generative systems to fan-facing explanations. The USTA editorial team oversees published content.

The IBM and USTA announcement and case study describe the oversight model.

This is applied AI inside an established product. Fans already use the app for scores and schedules. AI adds interpretation where the underlying data is useful.

The reach figures need care. More than 14 million people are expected to use the US Open’s digital properties in 2026, but that is projected audience size. It is not measured adoption of the new AI features.

The USTA and IBM should report feature usage, correction rates, editorial interventions and whether the explanations improve understanding or repeat engagement.

Control the workflow before expanding authority

These cases differ in industry and risk, but each system needs a defined operating boundary.

Product teams evaluating a similar opportunity should document six things before giving the system more authority:

The person accountable for the workflow
The data and systems the AI may access
The actions it may recommend or execute
The decisions that still require human approval
The customer or operating outcome expected to improve
The failure that would stop or reverse the rollout

Establish the current baseline for time, cost, errors, exceptions and user outcomes. Test the AI against the same cases.

Wider autonomy should follow evidence, not precede it.

Subscribe to Product Pulse Africa for practical analysis of AI, product development and financial services under African operating conditions.

Subscribe now

Further reading, sources and inspiration

Stripe: Agreement to acquire OpenRouter — the transaction and strategic rationale.
OpenRouter: Joining Stripe — current scale and what changes for users now.
AWS and Fanatics implementation account — architecture, escalation and reported support results.
Punch on Digital Encode’s DEPAS AI — the Nigerian launch, workflow and evidence gaps.
BusinessDay on DEPAS AI — the demonstration and company-reported deployment evidence.
OpenAI and Asana case study — the engineering migration process and reported cost.
IBM and USTA announcement — the new US Open features and data inputs.
IBM’s US Open case study — the product history and editorial oversight model.

Product Pulse Africa

Insight Scout | 25th August 2026