What changed
Up to 75% less manual prep per import
Case studies
AI & ML development
Thin Slice to Production with Acceptance Criteria
WORKFLOW:AI supplier price import with verify flow
SYSTEM STATE:Live pricing workflow, inconsistent supplier file formats
Proof in production
What we built
A clear checklist for what counts as ready, plus a rollout that added new formats one at a time.
What's protected
Nothing reaches production pricing without a human review first.
01
Get Energy moved supplier price imports into a controlled AI-assisted workflow
At Get Energy, supplier price sheets arrived in inconsistent formats and required manual preparation before import. The first production workflow was narrowed to one supported format, one bounded import flow, and one verification step before commit.
The main risks were branch mapping, period recognition, and price accuracy. Those risks stayed visible through human verification, staged format expansion, and clear ownership on both sides.
02
Context and constraints
Supplier files did not follow one predictable structure.
Different formats, branch variations, period layouts, and pricing representations made broad automation risky in the first release. A full-scope launch would have pushed parsing risk straight into a live commercial process.
What shaped the first release
Supplier files came in inconsistent structures
Commercial data had to stay correct before import
Manual preparation was slowing updates
The workflow needed human verification before commit
Format coverage had to grow in stages
03
Price imports were slowing commercial operations
At Get Energy, managers were spending time on repetitive preparation, manual checks, and coordination around price updates.
That delay affected internal alignment and slowed the path from updated pricing to signed contracts.
Why this workflow mattered
•Supplier files came in different structures
•Manual preparation slowed updates
•Branch mapping and period parsing carried direct commercial risk
•Delays in import slowed approval and contract flow
•Faster imports had a direct effect on operational speed and cash flow
04
The first release covered one narrow import path
The team started with one document format and one controlled path from upload to verified import. That kept the thin slice to production small enough to understand, test, and manage under live conditions.
More complex formats came later, after the first path was stable in use. That sequence kept scope realistic and reduced early blast radius.
What was inside the first slice
One supported document format
AI-assisted parsing of supplier price data
Branch mapping before import
Period and price extraction
Verify step before final import
Human confirmation before commit
What was added later
×Additional supplier-specific formats
×More complex matrix layouts
×Files without branches
×Files without scales
×Wider provider-specific prompt tuning
05
Data correctness defined the risk profile
The main risk was incorrect commercial data entering the workflow. A wrong import could send the wrong prices, periods, or branch mapping into the live process.
Correctness mattered more than speed alone. The first release had to protect the commercial flow before expanding automation coverage.
Highest-risk failure modes
01Incorrect branch mapping
02Wrong period recognition
03Wrong price extraction
04Structurally wrong output that still looked plausible
05False confidence in an import result that still needed review
06
Human verification reduced early launch risk
The system prepared the import result and exposed it before commit. That gave the team and the client a chance to catch mapping or parsing problems before they reached live pricing data.
The verify step made the first release safe enough to use while format coverage was still growing. It also created a clear boundary between AI assistance and irreversible operational change.
What was checked before confirmation
Supplier or branch interpretation
Period mapping to the expected range
Price placement in the correct fields
Final structure against the expected import shape
Missing or duplicated entries before import
07
Release confidence came from parsing stability across supported files
The workflow became releasable when supported file shapes could be parsed without structural errors and the output stayed clear enough for review.
The team widened support only after the current slice held up in live use.
Signs that the workflow was ready for release
•Supported formats parsed without structural breaks
•Mapping output remained readable and verifiable
•Import results were stable enough for real operational use
•New format support did not disturb the working path
•Coverage could expand gradually without losing control
08
Expansion happened format by format
The AI production rollout started with one supplier path. New formats and suppliers were added step by step, keeping early blast radius small and each addition contained.
That gave the team time to tune prompts, mapping logic, and verification behavior around real supplier documents. Coverage grew gradually until the workflow could handle the full operating set.
Expansion logic
Start with one supported format
Validate parsing and verification in live use
Add the next supplier or format after the previous path is stable
Extend prompt and mapping logic incrementally
Continue until coverage reaches the full operating set
09
Manual work dropped and pricing operations moved faster
The workflow cut manual prep time per import by up to 75%, removing repetitive preparation from the process. Coordination and approval around price updates became faster, moving commercial work forward with less delay.
The effect reached beyond convenience. Faster pricing updates supported quicker contract flow and healthier cash movement.
What improved
Less manual work for managers
Faster coordination around pricing updates
Quicker approval flow
Faster contract progression
Better operating rhythm around commercial changes
10
Ongoing quality depended on named owners on both sides
The workflow still needed active tuning after launch. Prompt behavior, mapping changes, and supplier-specific fixes remained part of the operating reality as coverage widened.
That work had clear owners on both teams. This kept the workflow maintainable as the import surface expanded.
Ownership boundaries
- Product Engineer on the delivery side owned prompt tuning and mapping behavior
- Commercial Director on the client side owned commercial fit and verification relevance
- Supplier-specific fixes were handled through continued tuning
- Workflow quality remained tied to review as new formats were added
One bounded workflow reached production through controlled expansion
This production AI case study shows how Get Energy turned one operational workflow into a first production slice, kept risk visible through verification, and expanded coverage step by step. That delivery shape fits workflows with clear operational value and a high cost of bad data.