Context engineering in a real workflow: which policy version applied to one service credit

Short answer

Context engineering is choosing exactly what information an AI system sees when it does a task: the instructions, tools, documents and history that matter, and nothing that misleads. In an ordinary business workflow it is the judgment an experienced employee makes before deciding, such as pulling the version of a service credit policy that was in force on the date of an outage, not the version in force today. A decision record that names the documents used captures a correct context.

Visual briefFull-size diagram
A timeline from April to July with two versions of a service credit schedule; the outage on May 28 falls under revision 3, while the request on June 3 and the approval on June 10 fall under revision 4, and the contract clause points back to the outage date.
Illustrative service credit at a fictional managed IT provider. Three dates sit on one ticket, and the contract decides which one selects the policy.

What does context engineering mean in an ordinary business workflow?

It means assembling the right inputs for one decision. In a business, that is the same work an experienced employee does when they pull the right contract, the right price list and the right policy version before acting.

Anthropic, which builds AI models, describes context engineering as the strategies for curating and maintaining the best set of information during a model's work, including instructions, tools, external data and message history. It frames the goal as finding the smallest set of high-signal information that makes the desired outcome most likely. More is not better. A model handed every policy revision ever written is as likely to quote the wrong one as a new hire would be.

That is why ordinary decision records are interesting. When a company logged which document a person relied on, it recorded a correct context selection. Our companion guide on context engineering and the documents behind a decision covers a credit limit case where the decision date selects the policy. The case below adds a harder twist: one ticket carries three dates, and a contract clause decides which one counts. For the broader idea of keeping records together rather than as isolated files, see how to identify useful workflow context.

One service credit, three dates (Illustrative)

The company is fictional: a 260-person managed IT services provider that owes customers service credits when an outage runs past a threshold in its service credit schedule. The schedule is a policy document kept in a versioned library.

  • Revision 3, in force January 1 to May 31: outages over 4 hours earn a credit of 15% of the monthly fee.
  • Revision 4, in force from June 1: the threshold drops to 2 hours, same 15%.
  • Master services agreement, clause 9.2: the schedule in force on the date of the incident governs.

A customer paying $9,600 a month suffers a 3-hour outage on May 28. They request a credit on June 3. The service director decides on June 10.

RecordWhat it saysSystemContext role
Incident ticketOutage May 28, 3 hours, root cause loggedTicketingSets the incident date and duration
Credit requestCustomer asks for an SLA credit, June 3Ticketing, linked to the incidentStarts the task
Clause 9.2Incident date selects the scheduleContract repositoryTells the decider which date counts
Schedule revision 34-hour thresholdDocument library, version historyThe rule that applies
Schedule revision 42-hour thresholdDocument library, current versionThe rule that looks right but does not apply
Billing recordMonthly fee $9,600Billing systemNeeded to compute any credit
Decision note"Rev 3 applies per 9.2; 3h below threshold; no SLA credit. Goodwill credit $480 approved."Ticketing, by service directorThe outcome and its stated basis

An assistant given only the current schedule would grant a credit of $1,440 (15% of $9,600) because 3 hours exceeds revision 4's threshold. An assistant given the clause and revision 3 reaches the director's answer: no contractual credit, plus a separate $480 goodwill gesture the director was authorized to make. The difference between those two answers is the whole point of context engineering, and the decision note is the evidence of which context was right.

The case also shows what to leave out. The request date and the decision date are both on the ticket and both fall under revision 4. Including them is fine; letting either one select the policy is the error. Good context is not only the right documents but the right reading of which detail governs.

The fields that make the right context recoverable

The decision can be rebuilt later only if a few fields survived alongside the ticket.

FieldWhere it sits in the exampleWhy the example depends on it
Incident date and durationIncident ticketThe facts the clause points to
Link between request and incidentTicketing relationStops the request date being mistaken for the incident date
Governing clause referenceContract repository, cited in the decision noteTells a reviewer which date selects the policy
Policy version with effective datesDocument library version historySharePoint, for example, keeps a version history from which earlier versions can be viewed and restored
Fee at the time of the incidentBilling recordAny credit is a share of this figure
Decision note naming the version usedTicket comment by the service directorRecords which context was chosen, in the decider's own words
Decider's authorityRole on the ticket or approval matrixShows the goodwill credit was within the director's limits

A model reviewer working from a tool such as micro1's flow, which lets a reviewer inspect a run's steps, tool calls and output against a rubric, would need exactly this kind of answer key to judge whether an agent picked the right context.

What the example shows, and what it cannot

  • Old versions may not survive. Microsoft notes that a SharePoint library keeps up to a set maximum number of versions and deletes the oldest when that limit is reached. A policy revised often can lose the version a past decision relied on.
  • A decision note is not always written. Here the director cited revision 3. Many decisions record only the outcome, leaving the context to be guessed.
  • One correct decision is not a rule. The goodwill credit was a judgment call. A different director might have refused it, and both could be within policy.
  • Access is not context. If the company lets AI assistants read its systems through the Model Context Protocol, that gives access to documents, not a record of which ones were right for a past decision, and it is not a license.

Who owns these records, and what stays unknown

For decision records like service credits, the owner is usually the COO or head of service delivery, with the CFO involved because credits touch revenue. Contracts sit with legal, and the document library with IT. Customer contracts may limit reuse of customer information, which is why rights review comes before any discussion of value.

Records only matter if the company can license them. Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English. The company also needs 50 or more full-time employees at peak (contractors excluded), documented operating history, rights to license the records, and an authorized contact.

What stays unknown until SourceX reviews it: how many decisions can be rebuilt with their full context, whether old versions survived, how personal and customer data will be handled, and whether any buyer wants this task type. Interest in context engineering across the industry does not mean every ticket archive qualifies. Partners never request exports, policy files or sample tickets. The owner can prepare with questions to ask before licensing company data. Customer support workflows covers the wider category, and more concepts are in the data types hub.

Go deeper on context engineering

Next step

Refer a company. No cap on rewards. Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.

If the service or finance lead confirms dated policy versions and decision notes that cite them, register as a partner, check the company against who qualifies, and read the terms before you make the introduction.

  1. Step 1Share your linkSend your personal link to a company you know.
  2. Step 2Company appliesThe company applies itself at /apply.
  3. Step 3Company signs and is paidThe company signs, delivers its data and is paid by SourceX under its agreement.
  4. Step 4You get your rewardthe greater of $50,000 plus 5% and 15% of the actual supplier payout, with no cap. A reward becomes payable after SourceX has paid the referred company under its signed agreement and the payout conditions in the program terms are met, including finance approval and your tax and payment details.

Common questions

Is context engineering the same as connecting an AI tool to company systems?

No. Connecting systems, for example through the Model Context Protocol, gives an assistant access. Context engineering is the choice of what the assistant should see for one task, and a licensing discussion is a separate question about rights to the records.

Does a company need an AI project of its own for this to matter?

No. What matters is whether its past decisions recorded which documents and rules were used. Many firms with no AI plans keep exactly that in their ticketing, billing and document systems.

Should I ask for the policy documents to check versions myself?

No. Ask whether old versions are kept, with effective dates, and in which system. SourceX reviews the documents themselves only under an agreement with the company.

Sources

  1. anthropic.com/engineering/effective-context-engineering-fo…
  2. micro1.ai/blog/introducing-flow
  3. support.microsoft.com/en-us/office/view-the-version-history-of-an-…
  4. support.microsoft.com/en-us/sharepoint/lists/documents-and-library…
  5. refer.sourcex.si/terms

Free resources

Know a company with valuable proprietary data?

Become a referral partner, get your link and introduce the company’s CEO (in an owner-run company, usually the owner). Referral partners can join from most countries. We can't work with partners in countries or regions under U.S. sanctions or on U.S. restricted-party lists. The companies you introduce should be based in the U.S., Canada and the EU, or in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia with at least 70% of their data in English.

Refer a company →

I own a business

Explore licensing your company's data directly to SourceX for AI training and evaluation. Start a short assessment; no uploads needed.

Start an assessment