Insight AI agents Automation AI adoption 22 September 2026

AI can now do real work. The useful decision is which work is worth handing to it

Alex Pletnev Founder and Managing Director, Inlight IT

AI can now do considerably more than draft an email or summarise a document. It can search across information, classify requests, monitor for changes, prepare decisions, update records and increasingly carry out steps across business systems. That expands the opportunity. It also makes the choice of work more important.

The useful question for a business is whether a piece of work is worth giving to AI once the licence, configuration, integration, checking, exceptions and cost of mistakes are included.

See the five tests

Work / value / verification / autonomy

Work Is this a task, or a whole piece of work?

Value Does it happen often enough to matter?

Verification Can the result be checked quickly?

Autonomy How much should AI be allowed to do?

Three questions

Questions that improve the AI conversation

Some work still belongs with a person, and structured, predictable work often belongs in conventional automation. These three questions sort out the rest, and they make a much better place to start than an AI feature list.

What makes a piece of work a good candidate for AI?

Look for work that happens frequently, carries enough value to justify changing it, involves language or variable information and produces an outcome that can be checked reasonably quickly. Document review, request triage, research, knowledge retrieval, exception handling and preparing a first decision are often more interesting candidates than a process where every step follows a fixed rule.

What it tests AI fit

When is ordinary automation still the better tool?

When the input is structured, the rules are stable and the same conditions should produce the same outcome every time. If an invoice below a threshold always follows the same approval path, or a completed form always creates the same record and notification, a workflow or integration can usually do that work reliably and inexpensively. Adding AI to a deterministic process can introduce cost and uncertainty without improving the outcome.

What it tests Tool choice

How much should AI actually be allowed to do?

Authority should follow consequence. AI can prepare, recommend and draft in areas where a person still owns the decision. It can execute within tighter boundaries when outcomes are easy to reverse, its performance is known and the business is comfortable with what it can access. The more a decision affects money, customers, employees, safety or regulatory obligations, the stronger the case for a human approval point.

What it tests Autonomy

Capability and value

AI capability has moved quickly while business value moves less evenly

The technology has crossed an important threshold. AI can increasingly work through a sequence of steps where it once answered one prompt at a time, and the major business platforms are embedding agents into the applications organisations already use. The results remain uneven, which makes task selection matter more.

What has moved

Works through multi-step sequences from a single request
Agents embedded in Microsoft, Salesforce, ServiceNow and other platforms
Cost model some capabilities licensed per user, others consumed by action, message or compute
Net effect easier to deploy, and easier to spend money on

What has not

Substantial experimentation with AI and agents across organisations
A much smaller group producing measurable financial value at scale
Emerging finding time saved by AI is partly consumed by checking, correcting and reworking its output
Implication the economic unit is the whole piece of work, including the checking

Licence

Configure

Integrate

Produce

Check

Handle exceptions

Finished work

A ten-minute task that takes eight minutes to check is a weak automation candidate. A three-hour research or document-handling process that can be reduced to twenty minutes of review is a different proposition.

Five tests

Five tests before you give AI the job

A useful AI use case usually survives five questions. Run the work through them before choosing the tool: the answers decide whether the job belongs with a person, in conventional automation, with assistive AI or with a bounded agent, and how much authority it should carry.

Test 01 · Does the work happen often enough to matter?

What to look for Frequency creates the economic base. A task that takes two hours but happens twice a year may irritate everyone involved without justifying a build. Twenty minutes repeated several hundred times a month is often more valuable.

In practice The useful measure is volume multiplied by the value of improving it. That value may be labour saved, faster turnaround, fewer errors, better service, better information or a bottleneck removed from someone whose time is more valuable elsewhere.

The rule Count how often the work happens, then multiply by what improving each instance is actually worth.

Test 02 · Is there enough variability for AI to add something?

What to look for Rules-based automation is exceptionally good at stable processes: when this happens, do that; if this field contains X, route it here; if the amount exceeds Y, request another approval. This is where the distinction from conventional automation becomes useful.

In practice AI earns more attention when rules become difficult to express because the input is language, documents, correspondence, images, inconsistent formats or judgement within a known boundary. A supplier invoice arriving in one of twenty layouts may benefit from AI extraction. The three-way match against purchase order, receipt and approved amount can still remain deterministic.

The rule Use each technology for the part of the problem it handles best.

Test 03 · Can someone verify the result quickly?

What to look for This may be the most important question. An AI-generated summary is useful because the source remains available and the reader can check anything important. An AI-prepared customer response can be reviewed before it is sent. A classification can be sampled against known examples.

In practice Verification becomes much harder when the AI completes a long chain of steps across several systems and the path to its outcome is hard to reconstruct. Where verification requires another person to redo the task from the beginning, much of the efficiency disappears.

The rule AI creates more value when producing the work is expensive and checking the work is cheap.

Test 04 · What happens when it gets something wrong?

What to look for A poor first draft of an internal project update is easy to correct. An incorrect payment, customer entitlement, employment decision or configuration change is different. The mistake may still be recoverable, but the cost of recovering it can be far higher than the saving created by automation.

In practice Think about reversibility as well as probability. A task can be suitable for AI even when the model is imperfect if errors are visible, inexpensive and easy to reverse. A task with rare but serious consequences deserves much tighter boundaries.

The rule The consequence of error should shape both the tool and the level of autonomy.

Test 05 · Is the process clear enough to automate?

What to look for AI handles variability better than traditional automation. It still needs a clear job. Someone needs to know what starts the work, what information matters, what an acceptable outcome looks like, which exceptions need escalation and who owns the result.

In practice Giving AI an unclear process usually produces an automated version of the ambiguity already there.

The rule The strongest candidates have enough structure to define the job and enough variability for AI to add value inside it.

The checking cost belongs in the business case. AI business cases often count the time removed from the first step and undercount the time added later: someone still reviews the draft, checks the source was interpreted correctly, takes over exceptions, unwinds failed actions and monitors the system. Workday’s January 2026 study of 3,200 employees, Beyond Productivity, found that roughly 40% of the time employees saved with AI was absorbed by rework: correcting errors, rewriting content and checking output. So the first useful metric is how long the completed, checked work took, compared with the original process. The same logic applies to consumption costs: agent platforms increasingly meter usage by action or workload, so a successful automation that runs thousands of times can become more expensive as adoption succeeds. That cost may be perfectly justified; it belongs in the calculation from the beginning.

Four responses

Four different types of work need four different responses

AI is one possible route for a piece of work, and the easiest mistake is treating it as the destination. Human-led work still makes sense when volume is low, context changes each time, relationships and negotiation matter or the consequences need a person to own the judgement. Stable, high-volume, rules-based processes remain excellent candidates for workflow automation, integration, Power Automate, RPA or application logic; approvals, notifications and predictable system updates rarely need an agent deciding what to do next.

Assistive AI is currently the broadest useful territory: research, summarise, extract, compare, classify, draft and prepare, with the person responsible for the outcome. Agents become interesting where the work is repeated, multi-step and spread across systems. The important word is bounded: a defined job, defined access, defined escalation points and a clear outcome.

Try it

Route a piece of work

Think of one recurring task and answer five questions. The page runs it through the tests above and tells you where it belongs.

01 How often does it happen?

02 What are the inputs?

03 Checking the result takes

04 If it gets it wrong

05 Is the job clearly defined?

01 Leave it with a person

02 Automate it conventionally

03 Use AI to assist the person

04 Give an agent a bounded job

Result

How much authority

This router runs on fixed rules, the same way a stable process should.

By function

The same department can contain all four

“Finance should use AI” or “customer service should use AI” is too broad to be useful. The work inside each function is different. Each one contains work that belongs with a person, work that suits conventional automation, work where assistive AI helps and a smaller amount that justifies a bounded agent.

Finance and people

Finance

A governed finance system should still produce the number. Conventional automation can handle matching, approvals and standard transaction workflows. AI can help interpret unstructured invoices, investigate anomalies, prepare management commentary or bring supporting information together for review. Payment authority and material financial decisions deserve a much higher bar.

HR

Policy questions, document retrieval, drafting, onboarding administration and routine request handling are natural areas to investigate. Recruitment, performance and termination decisions carry much greater consequence. AI may help organise the information; accountable people should remain close to the decision itself.

Customer-facing

Customer service

Order status, standard requests and information retrieval can be strong candidates for bounded automation or agents. Complaints, disputes, hardship and unusual cases involve context, empathy and consequence: AI can prepare the history and recommend a response while a person owns the interaction. High volumes can sit alongside weaker outcomes on complex interactions, so measure resolution quality as well as throughput. The useful model is often hybrid.

Sales

AI is already useful around the sale: account research, meeting preparation, proposal drafting, CRM hygiene and follow-up. The relationship, commercial judgement, negotiation and commitment remain human territory. A good implementation returns time to the salesperson and keeps automation out of the part customers actually value.

Operating the business

Operations and procurement

Document triage, supplier follow-up, scheduling, request classification and status chasing can all consume substantial time. These are especially interesting where the work is repetitive but not perfectly structured: the middle ground where ordinary workflow automation becomes cumbersome and AI can help interpret the variation.

IT support

Service desks already contain a natural progression. AI can summarise tickets and retrieve knowledge. It can then classify and recommend. With enough confidence, it can execute selected routine actions within a playbook. The level of authority should rise only as the action becomes well understood and recoverable.

Autonomy

Start with approval, then earn more autonomy

There is a useful progression between “AI suggested something” and “AI did it”, and most organisations do not need to jump to the end. A practical first deployment may allow AI to gather the information and prepare the action while a person clicks approve. That already removes a substantial amount of work.

Suggest

Prepare

Recommend

Execute with approval

Execute autonomously

An agent that saves five minutes but creates anxiety every time it acts has not solved much. Moving along this line is a business decision as much as a technical one.

More autonomy makes sense when

The task happens frequently and outcomes are measurable
Errors are easy to detect and inexpensive to reverse
Access is limited to what the task needs, and performance has been observed over time
The signal human approval is slowing the work more than it is protecting it

In Australia, consequential decisions

From 10 December 2026, APP 1.7 to 1.9 require an APP entity’s privacy policy to disclose where it arranges for a computer program to use personal information to make, or contribute substantially and directly to making, decisions that could reasonably be expected to significantly affect an individual’s rights or interests
Particularly relevant to employment, customer eligibility, access to services and other decisions involving personal information
The practical point as AI moves closer to the decision, understand what role it is actually playing: drafting information, or participating in a consequential decision

Our view

AI should earn its place in the workflow

The capability is moving quickly enough that almost every organisation will find useful applications for AI. The commercial discipline is choosing them well. The strongest use cases combine repeated value worth having, input varied enough for AI to handle, a result that can be verified efficiently and consequences that can be controlled. Stable rules still belong in stable automation. Human judgement still belongs with people. The objective is useful automation, with as much AI as the work justifies.

Route the work before choosing the tool

Measure the finished work, including checking

Use approval before increasing autonomy

Scale the use cases that survive the economics