Boost Blog

Trusting AI at Enterprise Scale

Written by Boost Payment Solutions | Nov 2026

What enterprise technology leaders need to rethink as AI moves from generating code to taking action

Written by Rinku Sharma, Chief Technology Officer at Boost Payment Solutions

We use AI enabled IDE’s, and there is no question these tools are changing software development. A capable engineer can get from an idea to working code much faster than was possible even a couple of years ago.

These IDE’s have changed the mechanics of day-to-day development. Some are very interesting to me for a different reason. Their spec-driven approach pushes requirements, design decisions and implementation tasks into the open before an agent starts building. That is useful because it forces more thinking upstream.

It also exposes a limitation that is easy to overlook: a well-structured specification can still be built on the wrong assumption.

If the underlying data model is misunderstood, or a business rule is implicit rather than documented, the agent may execute the specification exactly as written and still produce the wrong result. The tests may even pass because the tests were generated from the same assumption.

That is the part of AI-assisted engineering I find more interesting than the question of whether AI can write good code. Increasingly, it can. The issue is what happens when software production gets dramatically cheaper while the responsibility for operating that software does not.

A payment workflow makes the distinction obvious. Generating an endpoint that accepts a transaction is not particularly hard anymore. What takes more judgment is deciding how the system behaves when the processor accepts a payment but the response times out, whether a retry is safe, and how reconciliation works when two systems disagree about the transaction state.

AI can generate the implementation for all of those cases. Someone still needs to understand whether the assumptions behind that implementation are valid.

1. The Bottleneck is Shifting

Google's 2025 DORA research surveyed nearly 5,000 technology professionals. Ninety percent were already using AI at work and more than 80% said it improved productivity, while 30% reported little or no trust in AI-generated code. DORA also found that AI tends to amplify the engineering environment around it: organizations with stronger platforms and engineering practices generally get more value from it, while existing weaknesses can become more visible. [1]

That is consistent with what enterprise technology has always taught us. Faster execution helps enormously when the underlying system is sound. It is less helpful when the team is working around tightly coupled applications, stale documentation or business rules that exist differently in multiple places.
An experienced engineer will often stop at those points and ask why. An agent has to infer what the organization meant from the context it has been given.

Forrester's 2026 research describes software development moving beyond coding assistants toward agents that participate across planning, build, testing and delivery. [2] Kiro is one practical example of that progression because it moves work through requirements and design before implementation. [3]

I want engineering teams using these capabilities aggressively. The interesting problem now is making verification, context and authority scale at roughly the same pace as generation.

2. Engineering Changes. Judgment Becomes More Valuable

AI will affect engineering headcount and roles. An engineer who can accomplish significantly more with AI changes the economics of a technology organization, and companies will eventually redesign teams around that reality. Some activities will disappear entirely; others will require fewer people.
Where the simplistic version of that argument breaks down is in treating code production as the scarce capability.

Writing a retry function is straightforward. Knowing whether the business operation behind it is safe to retry can require understanding payment state, idempotency, reconciliation and how downstream systems behave when only part of a transaction succeeds.

The same distinction appears throughout engineering. AI can produce a database migration quickly, but somebody has to understand how it will behave against a production workload. It can modify Terraform, while deciding whether the resulting IAM policy creates an unacceptable security boundary requires a different kind of expertise. A generated test suite is useful only if the assumptions being tested are actually correct.

I expect strong engineers to spend less time producing syntax and more time defining systems, interrogating assumptions, designing verification and understanding the consequences of a change beyond the repository where it originated.

CodeRabbit's 2025 study provides an early data point worth considering. It examined 470 open-source pull requests and reported roughly 1.7 times as many identified issues in PRs it classified as AI-coauthored, with higher rates of logic, error-handling and security findings. The company also acknowledges that it inferred authorship and could not establish that every PR classified as human-only contained no AI assistance. [4]

That limitation matters, so I would not translate the research into a blanket claim that AI code is “1.7 times worse.”

The more practical conclusion is that increasing the amount of code entering the delivery pipeline without also improving review, testing and security does not automatically create more throughput. It can simply move the constraint downstream.

3. Data Quality Becomes an Operational Problem

The same shift is happening with enterprise data.

Imagine two companies using exactly the same frontier model. One has an authoritative customer definition, understood data lineage and clear ownership of important business metrics. The other has several definitions of revenue depending on whether the number comes from finance, sales or the data warehouse.

Both companies may have access to the same model. They do not have equivalent AI capabilities.

This used to be primarily an analytics problem. A bad definition might produce a questionable dashboard, and somebody would hopefully challenge the number before making a decision.

Agents change that dynamic. The result of a query can become input to another decision, which can trigger another system without waiting for a person to interpret the output manually.

DORA's AI Capabilities research identifies healthy data ecosystems as an important foundation for getting value from AI. [1] I think this becomes more consequential as autonomy increases.

Data quality, lineage and semantic consistency start to function as part of AI infrastructure. A better model can reason more effectively about the information it receives, but it cannot resolve a business definition that the organization itself has never settled.

There is an uncomfortable twist here as models improve: a wrong answer may become harder to detect because it is explained extremely well.

4. Agents Create a Different Security Problem

Traditional IAM is good at answering permission questions: which system a person can access, which API a service account can call, what an application is allowed to change.

Agents add another dimension because they can decide dynamically how to combine those permissions.
Consider an engineering agent with legitimate access to GitHub, Jira, Slack and part of the cloud environment. There may be a good reason for each permission individually. The security exposure comes from what the agent can accomplish by chaining them together.

Information retrieved from Slack could affect a code change in GitHub. If the same identity also has sufficient cloud permissions, that change could eventually alter infrastructure or reach production. The risk is not represented by any single permission; it sits in the sequence of actions the agent is capable of constructing.

That is the agentic blast radius security teams need to start modeling.

Netskope's 2026 AI report shows how rapidly this surface is expanding. In the telemetry it analyzed, remote MCP transactions increased 375% in ten weeks, and adoption of AI coding applications increased from 42% to 84% of organizations year over year. Netskope also identifies downstream data-policy violations and malicious-code risks as agents become connected to more enterprise systems. [5]

OWASP's 2026 Top 10 for Agentic Applications addresses issues such as goal hijacking, tool misuse, identity and privilege abuse, agentic supply-chain vulnerabilities and unexpected code execution. [6]

Those risks lead to a more nuanced security model than traditional least privilege alone. Agents need identities tied to purpose, limits around tool combinations, short-lived credentials where appropriate and explicit escalation before actions with a large blast radius.

There is no reason an agent that needs to inspect application logs should automatically inherit the authority to change production infrastructure. The same logic applies in financial systems: preparing a transaction and releasing funds are two very different permissions.

5. What I Would Build Around AI

Human approval cannot be the answer to every material agent action. If every autonomous workflow ends with somebody clicking a button, we will eventually constrain much of the value we were trying to create.
The model I prefer is risk-based autonomy.

Before an agent receives meaningful authority, the organization should know what outcome it is pursuing, which sources of enterprise information it should trust, what actions it can take and what evidence is required to accept the result. We should also be able to reconstruct what happened when something goes wrong.

Some of those controls should remain deterministic. Security scanning, transaction limits, environment boundaries and policy-as-code are not places where I want a model grading its own homework.

Other decisions depend on the consequences. A reversible action inside a development environment can tolerate much more autonomy than a change to production security, customer data or a financial transaction.

This is also where specifications help without becoming a substitute for judgment. Better requirements reduce ambiguity, but somebody still has to recognize what the requirements assumed and never explicitly stated.

NIST's Generative AI Profile takes a similarly broad approach, looking at AI risk across design, development, deployment, use and evaluation rather than treating the model itself as the entire problem. [7]

For me, governance is successful when it allows us to trust AI with more consequential work—not when it simply adds another approval process.

6. AI Changes the Build-Versus-Buy Calculation

The economics will eventually reach enterprise SaaS portfolios as well.

Some applications are mostly a workflow, a user interface and integrations around fairly well-understood business logic. As AI reduces the effort required to create and maintain those applications, capabilities that once made no economic sense to build internally may become viable.

That deserves a real reassessment of build versus buy, but it does not mean SaaS disappears.

When an enterprise buys a mature platform, it is also transferring part of the operational burden. Somebody else is responsible for keeping it secure, available and current, adapting it to regulatory changes and continuing to invest in the product.

Recreating the visible feature set is not the same as taking on those responsibilities.

I expect technology leaders to become more selective. Vendors that provide meaningful differentiated capability or absorb substantial operational complexity will continue to justify their economics. Thin workflow products will have a harder time defending the same premium once similar functionality becomes relatively inexpensive to build.

7. The leadership challenge is larger than model selection

A meaningful enterprise AI program eventually runs through architecture, data, cybersecurity and business operations. That is why I think the CAIO mandate, where companies choose to create the role, needs to extend well beyond choosing models or approving AI products.

The consequential decisions are about where machine intelligence belongs inside a business process, what context it receives, how much authority it has and what evidence is sufficient to trust the outcome.

AI expertise matters, but those decisions also require an understanding of system architecture, enterprise data, security and business risk. The technology itself is only one part of the operating model.

I am optimistic about where this leads because many of the benefits are already tangible. AI is taking tedious work out of development, helping engineers understand legacy systems and making internal software viable in situations where the economics would previously have killed the project before it started.

The next stage will be more demanding. As good models and capable coding agents become broadly available, having access to them will stop being much of a competitive advantage. What will matter is whether the company has given those systems something worth connecting to: data people actually trust, engineering foundations that can tolerate faster change, security controls designed for software that can act on its own, and enough discipline to recognize when automation should stop and judgment should take over.

AI is making software much easier to build, but the responsibility for what that software does has not moved anywhere. We still own the outcome.