How to Audit Your AI and SaaS Stack in 2026

A subscription list is the wrong map for a software audit. The map should show the work behind each tool, its full operating cost, the data it can reach, the actions it can take, and the evidence required before it is removed.

A general AI tool may cover several narrow drafting or classification tasks. It may not replace a system of record, a controlled approval process, reliable delivery, specialist analytics, or contractual support. Consolidation is successful only when the required work still passes its quality and risk thresholds after the old tool is gone.

Cancellation comes last. Inventory the stack, measure a representative baseline, run the replacement in parallel, test export and rollback, and then make one documented change.

Map the work behind each logo

A stack includes more than paid subscriptions. Record annual plans, usage-based services, free accounts, browser extensions, model APIs, automation connections, storage, and tools bought by individual team members. A service with a zero-dollar invoice can still create review work, expose data, or become difficult to leave.

Group tools by the job they support. These categories expose overlap without assuming that every product in a category is interchangeable:

The last two categories deserve a high threshold. Replacing a dedicated control with a conversational interface because both can produce a report is not functional equivalence.

Build one inventory with an owner for every tool

The NIST Cybersecurity Framework 2.0 includes inventories of software, services, systems, data, and supplier services in its asset-management outcomes. A small stack can use a spreadsheet. Larger organizations may already have procurement, identity, finance, and security records, but those sources still need to be reconciled.

Use one row per account, workspace, plan, or API project. Record:

• business owner and technical administrator;
• the job performed and the people who depend on it;
• fixed fee, seats, usage charges, add-ons, and renewal date;
• last meaningful use and the source that proves it;
• data categories stored, sent, or generated;
• authentication method, role, permissions, and connected systems;
• criticality and the consequence of an outage or incorrect action;
• available export formats, last export test, and import destination;
• contract owner, cancellation deadline, and deletion process.

An active login or invoice does not prove that a service still matters. Confirm the output the business needs, then look for shadow accounts and shared credentials. They may be missing from finance records but present in password managers, single sign-on logs, browser extensions, integration dashboards, or employee interviews.

Count the full operating cost

Use a consistent period and state every assumption. For a monthly comparison, allocate annual charges across 12 months and separate committed fees from variable usage.

monthly operating cost =
  fixed subscriptions and seats
  + usage, storage, add-ons, and transaction charges
  + administration and integration maintenance
  + review, correction, and exception handling
  + allocated implementation, migration, and exit work

This is a decision estimate, not an accounting standard. Use an agreed labor rate and measured time. Keep security, legal, continuity, and lock-in risk as separate decision fields when there is no defensible way to price them. Turning an uncertain risk into a precise dollar value only hides the uncertainty.

Calculate the current stack and the proposed stack with the same boundary. If an AI workflow makes the first draft faster but creates more review, retries, or maintenance, include those costs. Count a saving only after a fee or seat has actually been removed and the replacement has maintained the agreed service level for the observation period.

Find overlap with a capability matrix

Marketing feature lists are poor evidence of replacement. Build a matrix from representative work instead. Put required tasks in rows and tools in columns. For each cell, record one of four states: proven, partial, untested, or not supported. Link "proven" to a test result or an accepted production sample.

Break broad capabilities into testable units. "Content" might include source retrieval, outlining, factual review, brand approval, image rights, publishing, version history, and performance reporting. An AI assistant that drafts text covers one unit, not the whole process.

Then classify each tool:

• Keep: it performs a required job and no lower-risk option has passed the test.
• Downgrade: the job remains, but fewer seats, lower limits, or fewer add-ons are enough.
• Consolidate: another approved tool passes the same tasks and controls with lower total cost.
• Retire: the job is no longer required, or the tool duplicates a proven process.
• Investigate: ownership, usage, data, contract, or dependency evidence is missing.

When the evidence says "investigate," leave the decision open. Unknown data flows and forgotten automations are exactly what make a quick cancellation expensive.

Apply a data, permission, and resilience gate

Before testing a replacement, classify the data it will receive. Check contracts, client instructions, retention, deletion, training use, storage region, subprocessors, incident terms, and administrator visibility. The FTC's business guide to protecting personal information frames the work as a sequence: know what the company holds, keep only what it needs, protect the data, dispose of it safely, and prepare for incidents.

For a cloud service, compare the use case with the UK NCSC cloud security principles. They cover data protection, resilience, supply chain, identity and access, operational security, audit information, alerts, and secure defaults. A provider can be suitable for public drafting and unsuitable for confidential client records.

Limit every connection to the minimum data and authority required. A drafting pilot does not need permission to send messages. A classifier does not need delete access. A reporting assistant does not need payment authority. OWASP's guidance on sensitive information disclosure warns against relying on prompts as the main protection, while its guidance on excessive agency recommends narrow functionality, narrow permissions, downstream authorization, human approval for consequential actions, and monitoring.

Record a risk-gate result for each proposed replacement: pass, pass with controls, or fail. A lower invoice does not override a failed gate.

Run a parallel replacement test

Define acceptance before the pilot. Choose a fixed set of common cases, difficult cases, and failures. For AI-assisted work, include ambiguous input, missing fields, conflicting instructions, sensitive data, unsupported requests, malformed output, service timeouts, and instructions embedded in source material that should not be followed.

Measure at least:

• task success and factual or field accuracy;
• accepted, edited, rejected, and escalated results;
• review time and total time per accepted result;
• failed runs, duplicate actions, and recovery time;
• usage cost and maintenance time;
• permission, privacy, or policy exceptions.

Run the candidate beside the current process. Keep irreversible actions behind explicit approval and preserve a manual fallback. NIST publishes its AI Risk Management Framework Core as voluntary guidance. Evaluation begins before deployment and continues while a system is operating. Re-run the fixed set when the model, prompt, integration, permissions, or source data changes.

For a detailed method that starts with a low-risk draft and measures the complete workflow, see HUMAI's guide to AI automation for freelancers. The same baseline and review logic applies even when the operator is a team rather than a freelancer.

Prove the exit plan before cancellation

An export button is not an exit test. NIST's cloud security guidance treats termination planning and usable data export as part of managing cloud risk. Export a representative workspace, store it securely, and import it into the intended destination or a neutral viewer.

Check records, attachments, comments, versions, timestamps, identifiers, permissions, links, custom fields, automations, and audit history. Document what does not transfer and how it will be retained. Confirm the order for disabling integrations, revoking tokens, removing accounts, preserving required records, requesting deletion, and verifying the final invoice.

Name a rollback owner and a deadline. If the new process fails after cancellation, the team should know which data copy, manual procedure, and access path can restore service.

Keep a short decision record beside the inventory. It should name the service, proposed action, evidence reviewed, cost assumptions, acceptance threshold, risk-gate result, export-test result, approver, change date, and next review. This makes a later renewal easier and prevents a new owner from repeating the audit from memory. Record rejected consolidation ideas too, including the reason they failed.

Use one monthly audit cycle

1. Days 1 to 5: reconcile accounts, invoices, identity records, integrations, owners, and renewal dates.
2. Days 6 to 10: measure the baseline, calculate total operating cost, and build the capability matrix.
3. Days 11 to 20: test one candidate in parallel with fixed acceptance and risk criteria.
4. Days 21 to 25: test export, import, permission removal, fallback, and rollback.
5. Days 26 to 30: keep, downgrade, consolidate, retire, or investigate. Record the evidence and owner.

A monthly cycle is a planning device, not a promise that every contract can be changed in 30 days. Start before renewal and allow more time for seasonal workflows, regulated records, complex integrations, or annual commitments. Review high-change stacks on a defined cadence and trigger a new audit when price, model, terms, permissions, ownership, or data use changes.

The audit has done its job when every retained service has a current owner and a reason to exist, every proposed replacement has passed the same boundary of work, and the team can still recover if the change fails.