Module resource page supporting the session on agentic AI, the Spectrum of Agency, SOSTAC planning, and AI governance. Videos, frameworks, and key reading below.
Business analytics uses data, statistical methods, and technology to transform raw information into actionable insight. The four types form a maturity progression, each answering a different strategic question.
Summarises historical data to reveal patterns and trends. Dashboards, reports, and KPI tracking. The foundation all other analytics builds on.
Investigates root causes behind observed trends. Drill-down analysis, correlation studies, and anomaly detection to explain performance shifts.
Uses historical data, statistical models, and machine learning to forecast future outcomes. Regression, classification, and time-series forecasting.
Recommends optimal actions by analysing scenarios and constraints. Optimisation algorithms, simulation, and decision support systems. This is where agentic AI operates.
Figure 1.1 — The four types of business analytics form a maturity pyramid. Each level builds on the one below, and complexity (and strategic value) increases as you move upward. Agentic AI sits at the prescriptive apex because it both recommends and executes.
Organisations rarely skip levels. You cannot run prescriptive optimisation on data that has not been properly described, and you cannot diagnose a problem you have not measured. Most firms sit at the descriptive level: they have dashboards but cannot explain why metrics moved, let alone predict what will happen next. The investment required at each stage grows substantially. Descriptive analytics is largely off-the-shelf tooling (Power BI, Google Analytics). Diagnostic analytics requires analysts who can interrogate data. Predictive analytics requires statistical literacy and data engineering. Prescriptive analytics — and especially agentic AI on top of it — requires all of the above plus governance, data quality, and clear decision rights.
Agentic AI is not a replacement for the lower three levels. It is a layer on top of them. If your descriptive data is wrong, your prescriptive agent will make confidently wrong decisions at scale and at speed. The pyramid is a dependency stack, not a menu.
Agentic AI sits at the prescriptive end of this spectrum, but goes further: rather than recommending actions for a human to approve, it executes them autonomously. Understanding where your organisation sits on this analytics maturity curve is essential before deploying agentic systems.
Agentic AI systems autonomously plan, execute, evaluate, and iterate toward a goal — without waiting for human instruction at each step. They mark a shift from AI as a tool you prompt to AI as a system that pursues objectives with autonomy.
A systematic review confirms that agentic AI 'significantly improves productivity, reduces costs, and drives innovation' (Hosseini & Seilani, 2025). Understanding where and how to deploy these systems — and where to retain human oversight — is the central challenge for today's business analytics practitioners.
Figure 2.1 — The agentic loop. Unlike a chatbot that waits for your next prompt, an agent perceives its environment, reasons about what to do, acts, observes the result, reflects, and repeats. The "Act" node is highlighted because it is the step conventional AI cannot take without explicit human instruction.
An agent is more than a language model. It is a system composed of four components working together. The model is the reasoning engine (typically a large language model) that interprets instructions, plans, and generates responses. The tools are the actions the agent can take in the world: searching a database, sending an email, calling an API, updating a CRM record. The memory stores what the agent has learned within and across sessions, allowing it to build on prior reasoning rather than start from scratch. The orchestration layer is the loop shown above — it decides when to think, when to act, when to ask for help, and when to stop.
You can see the distinction clearly when you compare a chatbot and an agent given the same goal — say, "find me three case studies of brands using agentic AI and email me the links." A chatbot will draft an email you then have to send. An agent will search the web, evaluate sources, extract links, compose the email, and send it — pausing only if it encounters a constraint it cannot resolve.
Responds to instructions. You prompt, it answers, it waits. The human drives every step. Best for one-shot tasks: drafting, summarising, explaining.
Pursues goals with autonomy. It plans, acts, evaluates, iterates, and escalates when needed. Best for multi-step processes where the path is not fully known in advance.
SOSTAC (Smith, 1990; 2025 AI Edition) is one of the most widely adopted planning frameworks in business. Agentic AI fundamentally transforms how the TAC stages — Tactics, Action, Control — are executed.
Figure 3.1 — SOSTAC as a continuous planning loop. The SOS stages (Situation, Objectives, Strategy) require human judgement about context, ambition, and positioning. The TAC stages (Tactics, Action, Control) are operational execution — the layer where agentic AI delivers most value.
SOSTAC is a six-stage planning framework developed by PR Smith in 1990 and now in its sixth edition (with a 2025 AI edition). It works as a logical sequence: you cannot set objectives before understanding your situation, you cannot choose tactics before agreeing a strategy, and you cannot measure success before you have acted. Critically, the framework is also a loop — the Control stage feeds back into Situation Analysis for the next planning cycle, making it a continuous improvement model rather than a one-off plan.
| Stage | Component | Key Question | AI Impact |
|---|---|---|---|
| S | Situation Analysis | Where are we now? | AI-enhanced but human-interpreted |
| O | Objectives | Where do we want to be? | Human judgement required |
| S | Strategy | How do we get there? | Human creativity & positioning |
| T | Tactics | What is the detailed plan? | AI selects & optimises in real time |
| A | Action | How do we execute? | AI executes without manual intervention |
| C | Control | How do we measure? | AI creates self-correcting feedback loops |
Adapted from Smith (2025). SOS remains human-led; TAC is where agentic AI adds most value.
Many organisations misread SOSTAC as if AI should transform the whole model. It should not. The Situation, Objectives, and Strategy stages define what the business is for and how it competes — these are irreducible human judgements. The Tactics, Action, and Control stages define how the work gets done — and that is precisely where agents operate. Trying to delegate the SOS layer to AI is a category error; ignoring AI in the TAC layer leaves most of the value on the table.
Dave Chaffey's RACE framework (Reach, Act, Convert, Engage) provides a customer-centric structure for digital marketing planning. When combined with SOSTAC, it creates a complete decision architecture showing where agentic AI can automate the execution layer.
Figure 4.1 — The RACE funnel. Volume decreases at each stage, but customer value increases. The narrowest layer (Engage, marked with ⭐) is the most profitable over time, and it is also where agentic AI is safest to deploy: customers are known, data is rich, and decisions can be tested at small scale before being scaled.
RACE is a customer-lifecycle framework: it describes the journey a person takes from first hearing about a brand to becoming a loyal advocate. Each stage has a different goal, a different set of tactics, and — crucially — a different relationship with data and risk. This matters for agentic AI because not every stage of the funnel is equally suitable for automation.
The funnel narrows because each stage involves progressively more commitment from the customer. Reach involves a glance at a social post — low cost, low commitment, easy to replace. Engage involves a multi-year subscription or repeat purchase — high value, high switching cost, deeply personal. Agentic AI can be deployed at any stage, but the governance, monitoring, and human oversight required is most relaxed at the top of the funnel and most stringent at the bottom, where the consequences of a bad automated decision are largest.
| RACE Stage | Goal | Key Activities | Agentic AI Potential |
|---|---|---|---|
| Reach | Build awareness and traffic | SEO, paid search, social media, content marketing | AI optimises channel mix and bid strategies in real time |
| Act | Encourage interaction | Landing pages, content engagement, lead generation | AI personalises content and tests variations autonomously |
| Convert | Drive transactions | Conversion optimisation, retargeting, checkout flow | AI deploys hyper-targeted offers based on live behaviour |
| Engage | Build loyalty and advocacy | Email nurture, CRM, social advocacy, retention | AI monitors satisfaction, triggers retention campaigns, self-corrects |
Adapted from Chaffey and Ellis-Chadwick (2022) Digital Marketing: Strategy, Implementation and Practice. 8th edn.
Use SOSTAC to plan the overall marketing strategy (situation, objectives, strategy). Use RACE to structure the digital execution and customer journey (reach, act, convert, engage). When combined, SOSTAC provides the strategic SOS layer (human-led) and RACE provides the tactical execution layer where agentic AI operates most effectively.
Figure 4.2 — How SOSTAC and RACE fit together. SOSTAC sets the strategy; RACE structures the execution. The highlighted blue links show that the SOSTAC TAC layer (Tactics, Action, Control) directly powers the RACE execution stages (Act, Convert, Engage) where agentic AI operates.
A decision framework for positioning AI systems by their level of autonomy and the corresponding human role. Match the right level of independence to the risk, complexity, and governance requirements of each decision.
The right level depends on context. High-frequency, low-risk, data-rich decisions may benefit from full agency. High-stakes, reputationally sensitive decisions typically require human oversight (Ransbotham et al., 2025).
Figure 5.1 — The Spectrum of Agency is not a maturity ladder to climb. It is a menu of options to match to each decision. A pricing decision in a high-margin, data-rich context may be Fully Agentic; a price decision in a regulated, brand-sensitive context should remain Semi-Autonomous at most.
Most organisations make the mistake of treating the Spectrum as a single, organisation-wide choice — "we are an AI company" or "we are not." In practice, different decisions in the same organisation should sit at different points. The four questions that determine where any given decision should sit are: How reversible is the action? An email can be unsent; a price change is harder to walk back; a regulatory disclosure is hardest of all. What is the data quality? Agentic AI amplifies bad data into bad decisions at scale. What is the regulatory exposure? GDPR, the EU AI Act, and industry regulators impose hard constraints. What is the brand risk? A pricing error visible to a million customers is qualitatively different from a content moderation error seen by one.
Barney's (1991) Resource-Based View argues that sustained competitive advantage comes from resources that are Valuable, Rare, Inimitable, and Organised to capture value (VRIO). This lens explains why not all AI adoption creates competitive advantage.
Most strategic frameworks look outward: Porter's Five Forces examines industry structure, PESTLE scans the macro-environment, SWOT maps external threats and opportunities. The Resource-Based View (RBV), developed by Jay Barney in 1991, does something fundamentally different: it looks inward. It asks which internal resources your organisation possesses that competitors cannot easily replicate.
The practical application of RBV is the VRIO framework, which tests any resource against four sequential criteria. A resource must be Valuable (does it help exploit opportunities or neutralise threats?), Rare (do few competitors possess it?), Inimitable (is it costly for competitors to obtain or develop?), and the organisation must be Organised to exploit it (are the right processes, structures, and governance in place?). Failing any one of these tests downgrades the resource's strategic potential. Passing only Value produces competitive parity. Passing Value and Rarity produces temporary advantage. Passing all four produces sustained competitive advantage.
Figure 6.1 — The VRIO test as a sequential decision flow. Each "No" short-circuits the path to sustained advantage. A resource can be valuable and rare but still fail to deliver competitive advantage if competitors can imitate it or if the firm is not organised to exploit it.
The critical insight for business analytics practitioners is that AI tools exist on a spectrum from commodity to strategic asset, and most organisations are currently treating strategic technology as a commodity. A generic large language model like ChatGPT is valuable (it improves productivity), but it is not rare: every competitor has access to the same tool at the same price. Under VRIO, this produces competitive parity at best. As the European Journal of Innovation Management argues, 'competitive advantage does not stem solely from the availability or adoption of the technology itself, but rather from the firm's ability to integrate it with distinctive assets, such as proprietary data, specialised know-how, and an innovation-oriented organisational culture' (Emerald, 2025).
Agentic AI deployed on an organisation's own proprietary data changes this equation entirely. When an autonomous system is connected to data that only your organisation possesses, embedded within workflows that only your organisation operates, and governed by decision architectures that only your organisation has designed, it passes all four VRIO tests. The AI model itself is the commodity; your data and your organisational context are what make it a strategic resource.
| VRIO Test | Generic GPT / ChatGPT | Agentic AI on Proprietary Data |
|---|---|---|
| Valuable? | Yes - improves productivity | Yes - drives innovation, enables real-time decisions |
| Rare? | No - available to every competitor | Yes - built on unique organisational data |
| Inimitable? | No - any firm can adopt the same model | Yes - competitors cannot replicate your data or workflows |
| Organised? | Varies - often used ad hoc | Depends on governance design and decision architecture |
| Outcome | Competitive parity | Sustained competitive advantage |
Using ChatGPT to draft social media posts for Kellogg's gives no strategic edge, because Aldi, Lidl, and every cereal competitor can do exactly the same thing with the same tool at the same cost. It fails the Rarity test.
Deploying agentic AI connected to Kellogg's proprietary production line data, retailer sell-through analytics, and decades of seasonal demand patterns creates a resource competitors cannot replicate. The data is the moat, not the model.
The AI model is the commodity. Your data is the asset. Competitive advantage comes from connecting autonomous systems to proprietary data within governance structures you control (Barney, 1991; Emerald, 2025).
If an agentic system makes a pricing decision that causes reputational damage, who is accountable? As AI gains autonomy, traditional accountability structures break down.
Figure 7.1 — The governance stack for agentic AI. Trustworthy deployment is built in layers: data quality at the base, regulatory compliance near the top. Skipping any layer means a higher layer sits on a fragile foundation.
For decades, software governance was simple: a developer wrote code, the code was reviewed, the code was deployed, and the developer (or their employer) was accountable for what it did. Agentic AI breaks this chain in three ways. First, the system's behaviour emerges from the data it sees and the goals it is given, not from explicitly written code — meaning the "author" of any given decision is distributed and partly non-human. Second, the system can take actions in the world (sending emails, changing prices, contacting customers) that have real consequences, often before a human can intervene. Third, the same system can behave differently on different days as it learns and adapts, making pre-deployment testing insufficient.
The EU AI Act (Regulation 2024/1689), in force from August 2024 with phased application, responds to this by classifying AI systems into four risk tiers: unacceptable (banned outright — social scoring, manipulative subliminal techniques), high (regulated — credit scoring, recruitment, critical infrastructure), limited (transparency obligations — chatbots must disclose they are bots), and minimal (no specific obligations — spam filters, AI in video games). Agentic systems performing marketing decisions, customer service, or dynamic pricing will typically fall in the limited category, but can quickly slide into high if they affect access to services, employment, or fundamental rights.
Faster decisions, but no clear owner when things go wrong.
Real-time response at scale, but black-box reasoning and regulatory risk.
Reduced cognitive load on operators, but stakeholders may resist AI-led decisions.
Hyper-tailored outputs, but risk of algorithmic bias across demographics.
Before any agentic system is deployed in a real customer-facing context, three questions must be answered in writing: (1) What is the worst-case outcome if the agent malfunctions, and what is the rollback plan? (2) Who is the named human accountable for the agent's decisions, and do they have the authority to stop it? (3) Under which EU AI Act risk tier does this system fall, and what evidence do we have to support that classification? If these questions cannot be answered, the system is not ready for production.
Complete this assignment before the next session (Week 4: Designing Human-AI Decision Architectures). Bring your completed mapping to class for peer review.
Choose any organisation you are familiar with (your employer, a brand you use, or a case study from this module). Identify 3-5 key marketing or business decisions that the organisation currently makes. For each decision, complete the following:
Where does this decision currently sit on the Spectrum of Agency? (Reactive, Assistive, Semi-Autonomous, or Fully Agentic)
Where do you think it should sit, and why? Consider the data available, the risk level, and the speed of decision required.
Identify one governance, ethical, or privacy risk that would arise if this decision moved to full autonomy. Think about accountability, data protection, bias, or brand safety.
This exercise applies the Spectrum of Agency to a real context, preparing you for the Week 4 workshop on designing human-AI decision architectures. It also feeds directly into your summative assignment on Arriva Buses Wales.
To anchor the exercise, consider email send-time optimisation at a mid-sized retailer. Currently the marketing team chooses a fixed send window (e.g. Tuesday 10am). Under SOSTAC, this sits in the Tactics layer; under RACE, in the Act stage. The data is rich (open rates, click-through, conversion), the action is reversible (you can resend), and the consequences of a mistake are limited (a slightly worse open rate). The recommended position is Fully Agentic — let the agent choose the send time per recipient based on their individual engagement history. The governance risk to flag is consent creep: the system may begin sending outside the hours customers originally opted into, which is a GDPR compliance issue requiring an explicit update to the privacy notice and possibly a re-consent flow.