SAS 145 Risk Assessment: Templates vs Software vs AI
What changed in audit risk assessment, and why the documentation got heavier
Per the AICPA, SAS No. 145, Understanding the Entity and Its Environment and Assessing the Risks of Material Misstatement, superseded and amended AU-C section 315 and took effect for audits of financial statements for periods ending on or after December 15, 2023. Among other things it pushed firms toward separately assessing inherent risk and control risk, introduced inherent risk factors and a “spectrum” of inherent risk, and sharpened the required understanding of the entity’s IT environment and relevant general IT controls. Confirm the current text and any subsequent amendments directly in AU-C 315 — standards move, and this is one you don’t want to paraphrase from a blog post.
The practical consequence for a small or mid-size firm is more narrative. More why. More linkage between what you learned about the entity and the assertions you ended up testing. That’s exactly the kind of work that eats senior and manager hours without obviously producing audit evidence — which makes it a tempting automation target, and a dangerous one.
The three ways firms actually do this today
Templates and checklists. A Word or Excel risk assessment set, rolled forward each year by whoever has the engagement. Cheap, flexible, entirely dependent on the person filling it in. The failure mode is well known: last year’s memo gets a new date, and the “understanding of the entity” stops reflecting the entity.
Audit methodology software. Suites like CaseWare or Wolters Kluwer’s Knowledge Coach encode a methodology, link risks to procedures, and keep the documentation chain tight. They are genuinely good at structure and completeness — that’s the point of buying one. They are not good at reading your client’s general ledger and telling you what looks unusual this year. Several vendors have announced AI features; check the specific product documentation rather than the marketing page, and ask what data leaves your tenant.
An AI agent. Not a chatbot you paste a question into — an agent given scoped access to prior-year workpapers, the current trial balance and GL, and a defined job: assemble the inputs for risk assessment and produce a draft. Newest option, least standardized.
Where an agent genuinely helps on a risk assessment
Be specific, because “AI for audit” without specifics is how firms buy disappointment. The tasks that hold up:
- Year-over-year change detection. Pull the prior-year risk assessment and the current trial balance; list accounts with material movement, new accounts, new related parties, new revenue streams, and classes of transactions that didn’t exist last year. Assembly work with a checkable answer.
- Inquiry prep. Draft walkthrough and inquiry questions for each significant class of transactions, tailored to what changed rather than the generic list.
- IT environment inventory. Build a first-draft list of applications, hosting arrangements, and access-control owners from client responses and prior documentation, so the ITGC discussion starts from something.
- Candidate risk list with citations. Propose risks at the assertion level, each linked back to the document or journal entry population that prompted it. The citation requirement is what makes it reviewable.
- Gap-checking the memo. Have it compare the completed risk assessment against a documentation checklist you supply from your own methodology source — not from the model’s recollection of the standard, which may be stale, incomplete, or simply wrong. Feed it the checklist; let it flag what’s thin. A checker, not an author.
That last pattern — agent as reviewer rather than writer — is underrated. It pairs with the logic behind journal entry testing tools versus agents: a model that surfaces candidates for human attention is useful; a model that concludes is a liability.
An agent can assemble every input to a risk assessment and still have contributed zero audit evidence. The assessment is the judgment, and the judgment is the engagement.
What has to stay with a human
The assessment itself. Where a risk lands on the inherent risk spectrum, whether something is a significant risk, whether control risk is assessed at maximum, whether the planned response is sufficient — professional judgments the engagement team makes and documents. An agent proposing “significant risk” from pattern-matching last year’s file is producing a hypothesis, not a conclusion.
Professional skepticism is also not delegable. In practice, these models tend to accept a stated explanation unless explicitly prompted to challenge it — and even then, they can’t observe the pause before the answer. Humans sit in the room and follow up.
Automation, agents, and the difference that actually matters
Traditional accounting automation is deterministic: if the vendor is X, code to account Y. Rules are fast, cheap, auditable, and wrong only in predictable ways. An AI agent is probabilistic — it reads unstructured material, decides what to do next, and takes multi-step actions through tools you expose. It handles ambiguity rules can’t and fails in ways rules don’t, including confidently inventing a cross-reference.
For risk assessment, almost nothing is rule-automatable, because the inputs are narrative. That argues for an agent — inside a review gate, which is also why read-only MCP access is the right default before anyone discusses write permissions.
What the large firms run, and why copying it rarely works
Firms often ask what the Big 4 use. Broadly, the largest firms run proprietary audit platforms on commercial infrastructure, with internal data science teams and a methodology group controlling what the tooling is allowed to conclude. The transferable lesson isn’t the software — it’s the governance: a defined review gate, logging of what the tool produced versus what the human changed, and monitoring of output quality over time. That governance is implementable at small-firm scale. The platform is not, and shouldn’t be attempted.
A pilot you can run on one engagement
-
Pick one recurring client with a clean prior-year file
Not your most complex engagement. You want a case where you already know the right answer, so you can grade the output. -
Write the skill before you write the prompt
A skill is a reusable, packaged instruction set — your firm’s risk assessment procedure, required documentation elements, output format, and the rule that every proposed risk cites its source. Same discipline as building workpaper prep skills from a trial balance. -
Give it read-only access to a narrow slice
Prior-year risk assessment, current trial balance, GL detail, engagement letter, and correspondence for that engagement only. Scope access per engagement, log every call, keep write access off the table. -
Run it in parallel, not in place
The senior does risk assessment the normal way; the agent drafts alongside. Compare: what did it catch that the human missed, what did it miss, what did it fabricate? -
Decide based on the diff, then document the gate
If it earns a place, write down who reviews its output, what evidence of review goes in the file, and how a reviewer distinguishes agent-drafted text from human-authored conclusions. Your quality management monitoring should cover this.
Modeling the payoff without inventing a number
Skip the vendor ROI calculator and build your own:
(baseline hours per engagement − post-pilot hours) × number of engagements × fully loaded hourly cost
Pull the baseline from your time system for the risk assessment phase specifically, not the whole planning block. Then subtract costs that don’t appear in demos: writing and maintaining the skill, added manager review time, and subscription cost. The full picture includes recovered capacity — if planning hours come back in Q1, what gets reallocated to them, and is that work billable? Recovered hours that disappear into the week are not savings.
So will the CPA be automated away?
Pieces of the job already have been, and more will be. Risk assessment is a clean illustration of the boundary: the inputs automate well, the conclusion carries a signature, and signatures carry liability a model cannot hold. Standards assign responsibility to the engagement partner and the firm. Until that changes — and no standard-setter has signaled it will — AI shifts where auditors spend hours rather than eliminating the role.
The firms that get this right in 2026 won’t be the ones with the most AI features. They’ll be the ones that automated the assembly, kept the judgment, and can prove in a peer review which was which.
Not sure where to start?
Get a free automation audit: we map your bookkeeping, month-end close, client onboarding, document collection, and AP/AR — and show you what's worth automating before you spend a dollar.
Get a free automation audit