Signal

From company evidence to a clear, actionable AI transformation roadmap.

Comment

Inspiration

Small and medium-sized companies know where work is slow, repetitive, or frustrating, but they often do not have a technical team that can explain their systems, data, and AI readiness. Traditional digital-transformation consulting can also be too expensive or time-consuming for these businesses.

We built SIGNAL to give business leaders a practical starting point. Instead of asking them to understand technical architecture, SIGNAL begins with questions they can answer: Where does work happen? What causes delays? What outcome matters most? What evidence is available?

What it does

SIGNAL turns company context and selected evidence into a preliminary AI-transformation diagnostic.

The user describes the company, its operational problems, current systems, goals, leadership, skills, security, data, automation, and AI usage. They can optionally add a small number of redacted reports, policies, or system exports.

SIGNAL then produces:

  • A maturity profile across six fixed transformation dimensions
  • Separate evidence-strength and confidence labels
  • A comparison of possible workflow opportunities
  • Priority gaps connected to company evidence
  • Known unknowns, including who to ask and which records to request
  • A 90-day, six-month, and one-year roadmap
  • Named owners, measurable success signals, and stage gates
  • A downloadable Markdown or printable report

The result is intentionally presented as a preliminary decision-support tool—not an audit, certification, security clearance, or industry benchmark.

How we built it

We built SIGNAL with Next.js, TypeScript, Zod, the OpenAI Responses API, GPT-5.6, and OpenAI Codex.

GPT-5.6 analyzes messy business descriptions and selected evidence against a fixed, versioned rubric. It identifies a bounded workflow opportunity, evaluates readiness and dependencies, compares alternatives, and creates an evidence-linked transformation roadmap.

Application code—not the model—calculates the final maturity scores, evidence grades, confidence, coverage rules, and aggregation. Strict structured-output validation prevents the interface from receiving an unexpected report format.

The reasoning framework was informed by published digital-transformation and AI-deployment guidance from AWS, OpenAI, Meta, and the Anthropic/Goldman Sachs midmarket deployment initiative. These frameworks guide the recommendation but never count as evidence about the company.

Codex helped us design and implement the complete product: the evidence flow, scoring methodology, API integration, validation rules, report interface, exports, security controls, automated tests, and browser simulations.

Challenges we ran into

The hardest challenge was preventing the diagnostic from sounding more certain than its evidence justified.

Self-reported information can be useful, but it is not independently verified. We therefore separated maturity from evidence strength. Self-report remains E1 “Reported” with low confidence. A selected file can raise a relevant finding to E2 “Documented,” but the MVP does not claim high-confidence corroboration.

We also discovered that AI-generated reports could occasionally reference the wrong evidence ID, produce overly long text, or end a sentence at a length boundary. We addressed this with application-assigned evidence IDs, strict relationship checks, concise writing limits, script validation, and deterministic scoring.

Another challenge was balancing depth with usability. Business owners may not know their ERP architecture, so we replaced technical discovery questions with a plain-language system map and added known unknowns that explain exactly who to consult and what evidence to request.

Accomplishments that we're proud of

We are proud that SIGNAL is more than a generic chatbot. It combines AI interpretation with a fixed methodology and deterministic control layer.

Key accomplishments include:

  • Twenty-four fixed maturity lenses across six dimensions
  • Evidence citations behind decisions, gaps, and roadmap actions
  • Honest handling of missing information
  • Workflow comparisons instead of generic “use AI” recommendations
  • Location-specific privacy, labor, and regulatory questions
  • Numerical stage gates such as data-completeness and adoption thresholds
  • A deterministic sample mode for reliable demonstrations
  • Fifteen automated tests, successful type checking, and a production build
  • Live simulations using both self-report-only and documented-evidence scenarios

In one country-manager simulation, SIGNAL recommended a read-only replenishment pilot instead of autonomous purchasing. It required item-code mapping, human approval, representative pilot stores, and measurable inventory outcomes before expansion.

What we learned

We learned that a useful AI transformation recommendation depends less on producing a long list of AI ideas and more on connecting one business outcome to one bounded workflow.

We also learned that maturity, evidence strength, and confidence must remain separate. A company can describe an advanced capability while still having weak evidence for it. Likewise, missing information should result in an unknown or low-confidence finding—not an invented score.

Finally, we learned that small-business users do not need to become technology architects. A good diagnostic should translate their operational knowledge into the technical, data, governance, and security questions that specialists can answer later.

What's next for SIGNAL

The next step is to add secure, read-only integrations for sources such as Google Drive, SharePoint, ERP platforms, CRM systems, and accounting software. These integrations would use narrowly selected files and permissions rather than unrestricted company-wide access.

We also plan to add:

  • Progressive questionnaires based on industry and company size
  • Autosave and collaborative stakeholder input
  • Claim-level agreement and contradiction analysis
  • Expert-reviewed evaluation datasets
  • Side-by-side comparison of multiple transformation opportunities
  • Follow-up assessments that measure progress against earlier stage gates
  • Secure evidence drawers that show precisely how every conclusion was supported

Our goal is to make credible digital-transformation guidance more accessible while remaining transparent about what the AI knows, what it inferred, and what still needs to be verified.

Built With

Share this project:

Updates