CCAO-F 4.2 — Research, planning and process improvement

Which depth does the question need?

What does answering this actually require?

One or two lookups
Web searcha fact, a headline, a company detail
Hard thinking, no web
Extended thinkinganalysis over what you supplied
Many sources, synthesis
Researchfive or more searches, cited report
Answer lives in your files
Project or connectoryour documents, not the web

Anthropic’s guidance draws the line at roughly how many tool calls the question needs: one or two, none, or five and upward.

Research: what the feature actually does

Research is an agentic mode: rather than running one search and summarising the top results, Claude conducts multiple searches that build on each other, deciding what to investigate next as it goes, and returns an answer with citations the help centre describes as easy to check. Anthropic’s guidance positions it for questions needing roughly five or more tool calls and taking on the order of one to three minutes, and gives competitor comparisons, refreshing outdated internal documents, and triaging calendar and email actions as examples.

Two practical details matter for the exam and for your week. Research is available on the paid plans — Pro, Max, Team and Enterprise — and it requires web search to be enabled. And it uses your allowance faster than an ordinary chat: the help centre notes that research counts the same way against usage limits, but that a session can burn through them more quickly because Claude retrieves many sources. That is an argument for matching the tool to the question, not for avoiding the feature.

Research is not limited to the public web. When Google Workspace connectors are enabled, Claude can bring internal material — Gmail, Google Calendar, Google Docs — into the same investigation, so a question like “what did we promise this customer, and what has changed in the market since?” can be answered from both sides at once. Connectors mirror your existing permissions: Claude cannot reach anything you could not open yourself. Managing those connectors is covered in 5.2.

Question	Right depth	Why
“What is this supplier’s registered address?”	Web search	One lookup; a research run would be slower and no more correct
“Which of these three pricing options is least risky for us?”	Extended thinking	The inputs are yours; the work is reasoning, not retrieval
“How do our four main competitors price onboarding, and what changed this year?”	Research	Many sources, conflicting claims, needs synthesis and citations
“What did we agree in last month’s contract review?”	Project knowledge or a connector	The answer is in your documents, not on the web
“Draft a board paper on our AI pilot using the attached results”	Ordinary chat	Nothing to retrieve; it is a drafting task

A research request that returns something usable

Returns a readable essaytext

Research the market for procurement
software and tell me what you find.

Returns a decision inputtext

Research how mid-market procurement
software is priced, for a CFO buying
decision this quarter.

These four vendors only; sources from
the last 18 months only.

Per vendor: pricing model, typical
mid-market cost, extras, setup time.
Then a comparison table and the three
differences that matter to a 400-person
manufacturer.

Cite every figure. Where sources
disagree, show both and say which is
more recent. List what you could not
find rather than estimating it.

The strong brief fixes scope, recency, structure and — most importantly — what to do about disagreement and gaps.

The last instruction is the one people leave out. Without it, a gap in the evidence quietly becomes a confident sentence. With it, the gap arrives as a line you can act on — ask the vendor, or drop the criterion. Anthropic’s prompting guidance makes the general point: explain why the work matters and state the format you want, because current models follow instructions closely and do what was asked rather than volunteering extras.

Worked example · A competitive pricing scan before a renewal

Elena leads marketing operations at a software company. The renewal for their analytics vendor lands in six weeks, and the executive team wants to know whether the price is defensible.

She scopes it first. Four named competitors, pricing and packaging only, sources from the last 18 months, output as a table plus three bullet points for the exec summary.
She runs Research, not chat. The question needs many sources that must be reconciled — exactly what the mode is built for — and she leaves web search on so it can retrieve.
She reads the citations before the conclusions. Two of the price points come from a reseller’s blog rather than the vendor; one is three years old and marked as such. She drops both.
She checks the gaps section. Two vendors publish no public pricing. That becomes a question for the procurement team rather than an estimate in the deck.
She asks a follow-up in the same conversation for the implications for their own renewal — a reasoning task over material already gathered, with no further retrieval needed.

Takeaway: Research did the gathering and the reconciliation. Elena did the source triage, which took twenty minutes and removed two claims that would have been challenged in the room.

Planning: turning an outcome into a sequenced plan

Planning work has a reliable shape. You state the outcome and the date, list what must be true for it to happen, order those by dependency, attach owners, and then find the ways it could fail. Claude is fast at the middle three and genuinely useful at the last one, because listing failure modes is the step people are most likely to skip when they are already behind.

State the outcome, the date and the constraints — budget, headcount, anything fixed. A plan without constraints is a wish list.
Ask for a work-back plan from the date, with dependencies marked, rather than a forward list of tasks.
Name the owners yourself. Claude can suggest roles; only you know who actually has capacity.
Ask for the top risks and, for each, the early warning sign — the thing you would notice in week two if the plan were going wrong.
Ask what the plan assumes. Assumptions are where plans break, and they are invisible until they are listed.
Take the plan to the people doing the work before it becomes a commitment.

Process optimisation: measure before you redesign

Process work goes wrong in a predictable way: someone redesigns the process they imagine exists rather than the one people actually follow, and the new version fails at the same hidden step as the old one. The sequence that works is to map the current process from the people who run it, measure it, find where time and rework accumulate, redesign only those points, pilot, and measure the same thing again.

The improvement cycle

Map the real processsteps, handoffs, exceptions, who waits
Measure the baselinetime, volume, rework, error rate
Find the frictionqueues, rekeying, approvals, chasing
Redesign the worst stepone change at a time
Pilot and re-measuresame measure, same definition

Keep the measure; move to the next worst step

Claude helps most at the map and redesign steps. The measure steps are what make the redesign defensible — and they are what a scenario’s correct answer usually includes.

Claude earns its place at the map step because process knowledge is scattered across people’s heads and inboxes. Interview five people, paste the notes, and ask for a single step-by-step map with every handoff, every wait and every exception someone mentioned, marking where accounts differ. The disagreements are the finding: when two people describe the same approval differently, you have located either a training gap or an undocumented workaround.

Worked example · The weekly close report that ate ten hours

A finance manager at a services group produces a weekly forecast pack. It takes about ten hours across two people: pulling extracts, reconciling them, chasing three business units for commentary, and formatting the deck.

Map. Notes from both analysts go to Claude with a request for one process map showing each step, who waits on whom, and how long each step takes. The map shows that roughly half the elapsed time is chasing commentary, not analysis.
Measure. They record four weeks of actuals: hours per step, number of chase emails, number of late submissions.
Find the friction. The bottleneck is not the spreadsheet work. It is that business units submit free-text commentary in three different shapes, so every week is spent normalising it.
Redesign one step. A Project holds the template, the definitions and last week’s pack; commentary is pasted in and Claude returns it in the standard shape with figures cross-checked against the extract and anything unsupported flagged.
Pilot and re-measure. Same measure, four weeks. The chasing does not disappear, but the normalising work does.
Keep the human gate. The finance manager still reviews every figure before the pack is sent, because this is a financial document going to leadership.

Takeaway: The improvement came from measuring first. The obvious target was the spreadsheet work; the actual cost was inconsistent input formats, which nobody had counted. Anthropic’s published Advantage Solutions story describes the same pattern at scale — a weekly finance workflow of about ten hours compressed to under thirty minutes — but the reason it worked was that the process was understood before it was changed.

Redesign with and without a baseline

No baseline

“It feels faster” is the only evidence
The worst step was chosen by intuition
Several changes ship at once, so nothing is attributable
Savings claims cannot be defended when challenged

With a baseline

Hours, volume and rework counted before anything changed
The worst step was identified by the numbers
One change at a time, re-measured the same way
The claim survives scrutiny, and so does your credibility

The left column is how most process projects run. The right column is what makes an improvement claim survive a sceptical question in a steering meeting.

Traps the wrong answers are built from

Tempting but wrong	Do this instead
Using Research for a single factual lookup	Use web search when one or two lookups answer the question.
Asking a plain chat for current information	Turn on web search or Research; training data has a cutoff.
Accepting a research report without opening its citations	Check the sources behind any figure you will repeat in a meeting.
Redesigning a process from how you assume it works	Map it from the people who run it, and mark where their accounts differ.
Claiming a saving with no before-measure	Count hours, volume and rework before the change, and re-count the same way after.

You should now be able to

Choose between ordinary chat, web search, extended thinking and Research for a given question.
Write a research brief that fixes scope, recency, structure, citation and what to do about gaps.
Triage a cited research output before using its findings in a decision.
Produce a work-back plan with dependencies, owners, risks, early warning signs and stated assumptions.
Map and measure an existing process before redesigning any part of it.
Target a single high-friction step rather than automating a whole workflow at once.

Research, planning and process improvement